Category Archives: AI

Conway’s Law Is Dying

I’ve been thinking about Conway’s Law, the idea that organizations “ship their org chart.” The seams are most visible in big tech. Google, for example, once offered nearly a dozen messaging apps instead of a single excellent one, with each team fighting for resources. The same pattern appears everywhere: companies struggle to solve problems that cross organizational boundaries because bureaucracy and incentives keep everyone guarding their turf. The issue is not the technology; it is human nature.

I caught up with an old friend recently. We met at nineteen while working for one of the first cybersecurity companies, and now, in our fifties, we both advise organizations of every size on innovation and problem-solving. We agreed that defining the technical fix is the easy part; the hard part is steering it through people and politics. When change shows up, most organizations behave like an immune system attacking a foreign antibody. As Laurence J. Peter wrote in 1969, “Bureaucracy defends the status quo long past the time when the quo has lost its status.”

Naturally, the conversation drifted to AI and how it will, or will not, transform the companies we work with. I explored this in two recent posts [1,2]. We have seen the same thing: not everyone is good at using AI. The CSOs and CTOs we speak with struggle to help their teams use the technology well, while a handful of outliers become dramatically more productive. The gap is not access or budget; it is skill. Today’s AI rewards people who can break problems down, spot patterns, and think in systems. Treat the model like a coworker and you gain leverage; treat it like a tool and you barely notice a difference.

That leverage is even clearer for solo founders. A single entrepreneur can now stretch farther without venture money and sometimes never need it. With AI acting as marketer, product manager, developer, and support rep, one person can build and run products that once demanded whole teams. This loops back to Conway’s Law: when you are the entire org chart, the product stays coherent because there are no turf battles. Once layers of management appear, the seams show, and people ship their structure. Peter’s Principle follows, people rise to their level of incompetence, and the bureaucracy that emerges defends that status.

Yet while AI empowers outliers and small players, it might also entrench new kinds of monopolies. Big tech, with vast data and compute resources, could still dominate by outscaling everyone else, even if their org charts are messy. The question becomes whether organizational dysfunction will outweigh resource advantages, or whether sheer scale still wins despite structural problems.

The traditional buffers that let incumbents slumber (high engineering costs, feature arms races, and heavy compliance overhead) are eroding. Payroll keeps rising and headcount is the biggest line item, while the newest startups need fewer people every quarter. I expect a new wave of private-equity-style moves: smaller players snapped up, broken into leaner parts, and retooled around AI so they no longer rely on large teams.

Social media voices such as Codie Sanchez highlight the largest generational transfer of wealth in history. Many family-owned firms will soon be sold because their heirs have no interest in running them. These so-called boring businesses may look ripe for optimization, because most still rely on human capital to keep the lights on. Just above that tier we see larger enterprises weighed down by armies of people who perform repetitive tasks. A modern consulting firm armed with AI could walk into any of these firms and automate vast swaths of monotonous work that keeps those businesses running. Incumbents will struggle to move that fast, trapped by the very structures we have been discussing. A private-equity buyer, on the other hand, can apply the playbook with no sentimental ties and few political constraints.

ATMs let banks cut tellers and close branches. Customers later missed human service, so smaller neighborhood offices came back. AI will force every sector to strike its own balance between efficiency and relationship.

They say history doesn’t rhyme but it repeats, if so incumbents who dismiss AI as hype may follow Blockbuster into the museum of missed opportunities. In Wall Street (1987), Michael Douglas plays Gordon Gekko, a corporate raider who uses leveraged buyouts to seize firms like Blue Star Airlines, an aircraft maintenance and charter company. Gekko’s playbook, acquire, strip assets, slash jobs, was ruthless but effective, exploiting inefficiencies in bloated structures. Today, AI plays a similar role, not through buyouts but by enabling leaner, faster competitors to gut inefficiencies. Solo founders and AI-driven startups can now outpace large teams, while private-equity buyers use AI to retool acquired firms, automating repetitive tasks and shrinking headcounts. Just as Gekko hollowed out firms in any industry, AI’s relentless optimization threatens any business clinging to outdated, bureaucratic org charts.

Across news, television, music, and film, incumbents once clung to their near-monopoly positions and assumed they had time to adapt. Their unwillingness to face how the world was changing, and their instinct to defend the status quo, led to the same result: they failed to evolve and disappeared when the market moved on.

The Ask? incumbents, you need to automate before raiders do it for you.

What Does CPA Canada Have to Do With the WebPKI Anyway?

Leave a reply

When we discuss the WebPKI, we naturally focus on Certificate Authorities (CAs), browser root programs, and the standards established by the CA/Browser Forum. Yet for these standards to carry real weight, they must be translated into formal, auditable compliance regimes. This is where assurance frameworks enter the picture, typically building upon the foundational work of the CA/Browser Forum.

The WebTrust framework, overseen by professional accounting bodies, is only one way to translate CA/Browser Forum requirements into auditable criteria. In Europe, a parallel scheme relies on the European Telecommunications Standards Institute (ETSI) for the technical rules, with audits carried out by each country’s ISO/IEC 17065-accredited Conformity Assessment Bodies. Both frameworks follow the same pattern: they take the CA/Browser Forum standards and repackage them into structured compliance audit programs.

Understanding the power dynamics here is crucial. While these audits scrutinize CAs, they exercise no direct control over browser root programs. The root programs at Google, Apple, Microsoft, and Mozilla remain the ultimate arbiters. They maintain their own policies, standards, and processes that extend beyond what these audit regimes cover. No one compels the browsers to require WebTrust or ETSI audits; they volunteer because obtaining clean reports from auditors who have seen things in person helps them understand if the CA is competent and living up to their promises.

How WebTrust Actually Works

With this context established, let’s examine the WebTrust model prevalent across North America and other international jurisdictions. In North America, administration operates as a partnership between the AICPA (for the U.S.) and CPA Canada. For most other countries, CPA Canada directly manages international enrollment, collaborating with local accounting bodies like the HKICPA for professional oversight.

These organizations function through a defined sequence of procedural steps: First, they participate in the CA/Browser Forum to provide auditability perspectives. Second, they fork the core technical requirements and rebundle them as the WebTrust Principles and Criteria. Third, they license accounting firms to conduct audits based on these principles and criteria. Fourth, they oversee licensed practitioners through inspection and disciplinary processes.

The audit process follows a mechanical flow. CA management produces an Assertion Letter claiming compliance. The auditor then tests that assertion and produces an Attestation Report, a key data point for browser root programs. Upon successful completion, the CA can display the WebTrust seal.

This process creates a critical misconception about what the WebTrust seal actually signifies. Some marketing approaches position successful audits as a “gold seal” of approval, suggesting they represent the pinnacle of security and best practices. They do not. A clean WebTrust report simply confirms that a CA has met the bare minimum requirements for WebPKI participation, it represents the floor, not the ceiling. The danger emerges when CAs treat this floor as their target; these are often the same CAs responsible for significant mis-issuances and ultimate distrust by browser root programs.

Where Incentives Break Down

Does this system guarantee consistent, high-quality CA operations? The reality is that the system’s incentives and structure actively work against that goal. This isn’t a matter of malicious auditors; we’re dealing with human nature interacting with a flawed system, compounded by a critical gap between general audit principles and deep technical expertise.

Security professionals approach assessments expecting auditors to actively seek problems. That incentive doesn’t exist here. CPA audits are fundamentally designed for financial compliance verification, ensuring documented procedures match stated policies. Security assessments, by contrast, actively hunt for vulnerabilities and weaknesses. These represent entirely different audit philosophies: one seeks to confirm documented compliance, the other seeks to discover hidden risks.

This philosophical gap becomes critical when deep technical expertise meets general accounting principles. Even with impeccably ethical and principled auditors, you can’t catch what you don’t understand. A financial auditor trained to verify that procedures are documented and followed may completely miss that a technically sound procedure creates serious security vulnerabilities.

This creates a two-layer problem. First, subtle but critical ambiguities or absent content in a CA’s Certification Practice Statement (CPS) and practices might not register as problems to non-specialists. Second, even when auditors do spot vague language, commercial pressures create an impossible dilemma: push the customer toward greater specificity (risking the engagement and future revenue), or let it slide due to the absence of explicit requirements.

This dynamic creates a classic moral hazard, an issue similar to the one we explored in our recent post, Auditors are paid by the very entities they’re supposed to scrutinize critically, creating incentives to overlook issues in order to maintain business relationships. Meanwhile, the consequences of missed problems, security failures, compromised trust, and operational disruptions fall on the broader WebPKI ecosystem and billions of relying parties who had no voice in the audit process. This dynamic drives the inconsistencies we observe today and reflects a broader moral hazard problem plaguing the entire WebPKI ecosystem, where those making critical security decisions rarely bear the full consequences of poor choices.

This reality presents a prime opportunity for disruption through intelligent automation. The core problem lies in expertise “illiquidity”, deep compliance knowledge remains locked in specialists’ minds, trapped in manual processes, and is prohibitively expensive to scale.

Current compliance automation has only created “automation asymmetry,” empowering auditees to generate voluminous, polished artifacts that overwhelm manual auditors. This transforms audits from operational fact-finding into reviews of well-presented fiction.

The solution requires creating true “skill liquidity” through AI: not just another LLM, but an intelligent compliance platform embedding structured knowledge from seasoned experts. This system would feature an ontology of controls, evidence requirements, and policy interdependencies, capable of performing the brutally time-consuming rote work that consumes up to 30% of manual audits: policy mapping, change log scrutiny, with superior speed and consistency.

When auditors and program administrators gain access to this capability, the incentive model fundamentally transforms. AI can objectively flag ambiguities and baseline deviations that humans might feel pressured to overlook or lack the skill to notice, directly addressing the moral hazard inherent in the current system. When compliance findings become objective data points generated by intelligent systems rather than subjective judgments influenced by commercial relationships, they become much harder to ignore or rationalize away.

This transformation liquefies rote work, liberating human experts to focus on what truly matters: making high-stakes judgment calls, investigating system-flagged anomalies, and assessing control effectiveness rather than mere documented existence. This elevation transforms auditors from box-checkers into genuine strategic advisors, addressing the system’s core ethical challenges.

This new transparency and accountability shifts the entire dynamic. Audited entities can evolve from reactive fire drills to proactive, continuous self-assurance. Auditors, with amplified expertise and judgment focused on true anomalies rather than ambiguous documentation, can deliver exponentially greater value.

Moving Past the Performance

This brings us back to the fundamental issue: the biggest problem in communication is the illusion that it has occurred. Today’s use of the word “audit” creates a dangerous illusion of deep security assessment.

By leveraging AI to create skill liquidity, we can finally move past this illusion by automating the more mundane audit elements giving space where the assumed security and correctness assessments also happen. We can forge a future where compliance transcends audit performance theater, becoming instead a foundation of verifiable, continuous operational integrity, built on truly accessible expertise rather than scarce, locked-away knowledge.

The WebPKI ecosystem deserves better than the bare minimum. With the right tools and transformed incentives, we can finally deliver it.

From Mandate to Maybe: The Quiet Unwinding of Federal Cybersecurity Policy

Leave a reply

Why the 2025 Amendments to EO 14144 Walked Back Progress on PQC, SBOMs, and Enforcement, Even as the Products to Support Them Have Become Real.

The June 2025 amendments to Executive Order 14144 read like a cybersecurity manifesto. They name adversaries (China, Russia, Iran, North Korea) with unprecedented directness and reference cutting-edge threats like quantum computing and AI-enabled attacks. The rhetoric is strong. The tone, urgent.

But beneath the geopolitical theater, something quieter and more troubling has happened. The Executive Order has systematically stripped out the enforcement mechanisms that made federal cybersecurity modernization possible. Mandates have become “guidance.” Deadlines have turned into discretion. Requirements have transformed into recommendations.

We’re witnessing a shift from actionable federal cybersecurity policy to a fragmented, voluntary approach, just as other nations double down on binding standards and enforcement.

The Enforcement Rollback

The most visible casualty was the software bill of materials (SBOM) mandate . The original EO 14144 required vendors to submit machine-readable attestations, with specific deadlines for updating federal procurement rules. These requirements have been entirely deleted.

This removal actually makes sense. Most SBOMs today are fundamentally broken: generated manually, and don’t actually match to deployed artifacts. Without robust validation infrastructure, SBOMs create more noise than signal. Use cases like vulnerability correlation break down when the underlying data is untrustworthy.

Once you have reproducible builds and verifiable provenance pipelines, SBOMs become implicit in the process. The government was both premature and naive in requiring SBOMs before the ecosystem could reliably generate them and do something with them. More fundamentally, they hooed that mandating documentation would somehow solve the underlying supply chain visibility problem – unfortunately thats not the case.

But SBOMs are a symptom of deeper issues: unreproducible builds, opaque dependency management, and post-hoc artifact tracking. Simply requiring vendors to produce better paperwork was never going to address these foundational challenges. The mandate confused the deliverable with the capability.

What’s more concerning is what else disappeared. Provisions mandating phishing-resistant multi-factor authentication, real-time interagency threat sharing, and specific timelines for aligning federal IT procurement with Zero Trust requirements all vanished. The detailed Border Gateway Protocol security language was replaced with generic “agency coordination” directives. The EO stripped away near-term pressure on vendors and agencies alike.

Yet even as these enforcement mechanisms were being removed, the amendments introduced something potentially transformative.

Rules as Code: Promise, Paradox, and Perfect Timing

The most exciting addition is buried in bureaucratic language. A pilot program for “machine-readable versions of policy and guidance” in cybersecurity appears almost as an afterthought. While the EO doesn’t name OSCAL explicitly, this is almost certainly referring to expanding the Open Security Controls Assessment Language use beyond its current FedRAMP usage into broader policy areas.

This could be transformative. Imagine cybersecurity policies that are automatically testable, compliance that’s continuously verifiable, and security controls that integrate directly with infrastructure-as-code. OSCAL has already proven this works in FedRAMP: structured security plans, automated assessment results, and machine-readable control catalogs. Expanding this approach could revolutionize how government manages cybersecurity risk.

But there’s something deliciously ironic about the timing. We’re finally standardizing JSON schemas for control matrices and policy frameworks just as AI becomes sophisticated enough to parse and understand unstructured policy documents directly. It’s almost comical. Decades of manual compliance work have driven us to create machine-readable standards, and now we have “magical AI” that could theoretically read the original messy documents.

Yet the structured approach remains the right direction. While AI can parse natural language policies, it introduces interpretation variations. Different models might understand the same requirement slightly differently. OSCAL’s structured format eliminates ambiguity. When a control is defined in JSON, there’s no room for misinterpretation about implementation requirements.

More importantly, having machine-readable controls means compliance tools, security scanners, and infrastructure-as-code pipelines can directly consume and act on requirements without any parsing layer. The automation becomes more reliable and faster than AI interpretation. Real-time compliance monitoring really only works with structured data. AI might tell you what a policy says, but OSCAL helps you build systems that automatically check if you’re meeting it continuously.

This pattern of promising technical advancement while retreating from enforcement continues in the amendments’ approach to cryptographic modernization.

The Post-Quantum Reality Check

Then there’s the post-quantum cryptography provisions. The EO requires CISA and NSA to publish lists of PQC-supporting products by December 2025, and mandates TLS 1.3 by January 2030.

The TLS 1.3 requirement appears to be carried over from the previous executive order, suggesting this wasn’t a deliberate policy decision but administrative continuity. The amendment specifically states that agencies must “support, as soon as practicable, but not later than January 2, 2030, Transport Layer Security protocol version 1.3 or a successor version.” More tellingly, the 2030 timeline likely reflects a sobering recognition of enforcement reality: federal agencies and contractors are struggling with basic infrastructure modernization, making even a five-year runway for TLS 1.3 adoption potentially optimistic.

This reveals the central tension in federal cybersecurity policy. The infrastructure is calcified. Legacy systems, interception-dependent security architectures, and procurement cycles that move at geological speed all contribute to the problem. A 2030 TLS 1.3 mandate isn’t visionary; it’s an acknowledgment that the federal government can’t move faster than its most outdated components.

But this enforcement realism makes the broader PQC timeline even more concerning. If we need five years to achieve TLS 1.3 adoption across federal systems, how long will the actual post-quantum migration take? By 2030, the question won’t be whether agencies support TLS 1.3, but whether they’ve successfully migrated key exchange, digital signatures, and PKI infrastructure to post-quantum algorithms. That’s a far more complex undertaking.

In essence, the EO treats PQC like a checklist item when it’s actually a teardown and rebuild of our cryptographic foundation. Historically, the federal government has led cryptographic transitions by creating market demand and demonstrating feasibility, not by setting distant mandates. When the government moved to AES or adopted Suite B algorithms, it drove adoption through procurement pressure and early implementation.

Meanwhile, allies like the UK and Germany are taking this traditional approach with PQC. The UK’s National Cyber Security Centre has published detailed migration timelines and will launch a pilot program to certify consultancy firms that provide PQC migration support to organizations. Germany’s Federal Office for Information Security has been leading in co-developing standards and demonstrating early government adoption. They’re creating market pull through demonstrated feasibility, not regulatory deadlines that may prove unenforceable.

Beyond cryptography, the EO does introduce some concrete requirements, though these represent a mixed bag of genuine progress and missed opportunities.

The EO also tasks NIST with updating key frameworks and calls for AI-specific vulnerability coordination. All valuable work. But notably absent: any requirement for agencies to adopt, implement, or report on these updated frameworks.

One genuinely new addition is the IoT Cyber Trust Mark requirement: by January 2027, federal agencies must require vendors of consumer IoT products to carry the labeling. This represents concrete procurement leverage, though it’s limited to a narrow product category.

These mixed signals, technical infrastructure development alongside enforcement retreat, reflect a broader pattern that undermines the federal government’s cybersecurity leadership.

As we’ve explored in previous discussions of AI’s impact on compliance, this shift toward automated policy interpretation and enforcement represents a broader transformation in how expertise flows through complex systems, but only when the underlying mandates exist to make that automation meaningful.

We’re building this sophisticated machine-readable infrastructure just as the enforcement mechanisms that would make it meaningful are being stripped away. It’s like having a perfectly engineered sports car but removing the requirement to actually drive anywhere.

The Infrastructure Is Ready. The Mandate Isn’t.

Federal cybersecurity policy shapes vendor behavior, influences state and local government standards, and signals U.S. priorities to international partners. Without centralized mandates, vendors receive mixed signals. Agencies implement inconsistently. Meanwhile, international partners advance with clearer timelines and stronger enforcement. The U.S. risks ceding leadership in areas where it built the foundational standards, just as adversaries accelerate their own capabilities.

The United States has built remarkable cybersecurity infrastructure. OSCAL for automated compliance, frameworks for secure software development, and draft PQC standards for cryptographic transition all represent genuine achievements. But the June 2025 amendments represent a retreat from the leadership needed to activate this infrastructure.

We have the tooling, standards, and momentum, but we’ve paused at the moment we needed to press forward. In the face of growing threats and global urgency, discretion is not resilience.

We’ve codified trust, but stopped requiring it, leaving security to agency discretion instead of institutional design. That’s not a strategy. It’s a hope. And hope is not a security control.

Rethinking Compliance: AI, Skill Liquidity, and the Quest for Verifiable Truth

2 Replies

In an earlier piece, ‘The Limitations of Audits,’ we explored how traditional compliance frameworks often fall short, functioning as point-in-time assessments rather than drivers of continuous security practices. Building on that foundation, and expanding on our exploration in ‘When AI Injects Liquidity Into Skills: What Happens to the Middle Tier?’, let’s examine how AI is poised to transform this landscape by introducing “skill liquidity” to compliance and auditing.

The High Price of Illiquid Expertise: Manual Bottlenecks in Compliance Today

As I’ve lamented before, the real cost of traditional, “illiquid” approaches to compliance expertise is staggering. In WebTrust audits, for instance, audit teams frequently report not having “enough time to look at the big picture” because their efforts are consumed by manual, repetitive tasks. Approximately 5-10% of an entire audit engagement – which can range from 350 to well over 1,500 hours for the audit firm alone – is often dedicated just to mapping organizational policy documents against standard templates. Another 15-20% of those hours are spent scrutinizing core operational processes mandated by frameworks, such as user access lifecycles or system change logs.

These percentages represent an enormous drain of highly skilled human capital on work that is largely automatable. And these figures only account for the auditors’ direct engagement. The true cost multiplies when you factor in the mountain of preparation by the entity being audited and subsequent review by third parties. The fully loaded headcount costs across this ecosystem for a single audit cycle represent a heavy tax on expertise that remains stubbornly “frozen” in manual processes.

First-Wave Automation: A Trickle of Skill Liquidity, or a New Kind of Friction?

The first wave of automation has arrived, with tools like Vanta and Secureframe offering streamlined pathways to certifications like SOC 2 by generating policy templates and automating some evidence collection. For many organizations, especially those with simpler, cloud-native environments, this has made basic compliance more accessible, a welcome “trickle of skill liquidity” that helps get a generic certification done in record time.

However, this initial wave has inadvertently created what we might call “automation asymmetry.” These tools predominantly empower the audited entity. When a company uses sophisticated automation to produce voluminous, perfectly formatted artifacts, while auditors still rely on largely manual review, a dangerous gap emerges. The truth risks getting lost in these “polished milquetoast” audits. The sheer volume and veneer of perfection can overwhelm human scrutiny, potentially masking underlying issues or a compliance posture that’s merely superficial. The audit can devolve into a review of well-presented fiction rather than an unearthing of operational fact.

Unlocking True Skill Liquidity: Intelligent Systems That Make Deep Compliance Knowledge Flow

To move beyond surface-level automation or basic Large Language Models (LLMs), we need intelligent compliance systems – sophisticated platforms designed to embed and scale deep domain knowledge. This isn’t just about processing text; it’s about an AI that understands context, relationships, history, and the intricate rules of specific compliance frameworks from the perspective of all stakeholders. Indeed, this drive to embed and scale specialized knowledge through AI is a significant trend across industries. For instance, leading professional services firms have been developing proprietary generative AI platforms, like McKinsey’s Lilli (announced in 2023), to provide their consultants with rapid access to synthesized insights drawn from vast internal knowledge bases, effectively enhancing their own ‘skill liquidity’ and analytical capabilities. Such systems, whether for broad consulting or specialized compliance, require:

An ontology of expertise: Encoding the structured knowledge of seasoned auditors—controls, their intent, interdependencies, and valid evidence criteria.
An ontology of documents: Understanding the purpose and interplay of diverse artifacts like System Security Plans, policies, vulnerability scans, and their connection to the compliance narrative.
Temporal logic and change tracking: Recognizing that compliance is dynamic, and analyzing how policies, controls, and evidence evolve over time, identifying drift from baselines.
Systemic integration: A cohesive architecture of LLMs, knowledge graphs, rule engines, and data connectors that can ingest, analyze, and provide auditable insights.

This approach transforms an AI from one that simply helps prepare artifacts to one that can critically assess them with genuine understanding – a crucial shift towards making knowledge truly usable (a concept we delve into in ‘From Plato to AI: Why Understanding Matters More Than Information’ ) – making that deep compliance knowledge flow across the ecosystem.

Liquidating Rote Work, Elevating Human Expertise: AI’s Impact on Audit Value and Integrity

When auditors and program administrators leverage intelligent systems, the nature of their work fundamentally changes—a direct consequence of “skill liquidity.” The AI can ingest and critically analyze the (potentially voluminous and auditee-generated) artifacts, performing the initial, labor-intensive review that consumes so many hours. This liquidates the rote work, significantly impacting even the global delivery models of audit services, as routine document review tasks are often offshored for cost savings, can now be performed with greater consistency, speed, and contextual insight by these intelligent systems.

This frees up high-value human experts to:

Focus on what truly matters: Shift from the minutiae of “collection, ticketing, whether there was testing involved, whether there was sign-off” to the crucial judgment calls: “Is this a finding or a recommendation?”
Investigate with depth: Dive into complex system interactions, probe anomalies flagged by the AI, and assess the effectiveness of controls, not just their documented existence.
Enhance audit integrity: By piercing the veneer of “polished” evidence, these AI-augmented auditors can ensure a more thorough and truthful assessment, upholding the value of the audit itself.

The New Compliance Economy: How Liquid Skills Reshape Teams, Tools, and Trust

This widespread skill liquidity will inevitably reshape the “compliance economy.” We’ll see:

Transformed Team Structures: Fewer people will be needed for the easily automated, “liquid” tasks of data collection and basic checking. The demand will surge for deep subject matter experts who can design, oversee, and interpret the findings of these intelligent systems, and who can tackle the complex strategic issues that AI surfaces.
Empowered Audited Organizations: Companies won’t just be scrambling for periodic audits. They’ll leverage their own intelligent systems for continuous self-assurance, drastically reducing acute audit preparation pain and eliminating those “last-minute surprises.” Furthermore, the common issue of “accepted risks” or Plans of Action & Milestones (POA&Ms) languishing indefinitely is addressed when intelligent systems continuously track their status, aging, and evidence of progress, bringing persistent, transparent visibility to unresolved issues.
New Proactive Capabilities: With compliance intelligence more readily available, organizations can embed it directly into their operations. Imagine Infrastructure as Code (IaC) being automatically validated against security policies before deployment, or proposed system changes being instantly assessed for policy impact. This is proactive compliance, fueled by accessible expertise.

Trust is enhanced because the processes become more transparent, continuous, and validated with a depth previously unachievable at scale.

The Liquid Future: Verifiable, Continuous Assurance Built on Accessible Expertise

The ultimate promise of AI-driven skill liquidity in compliance is a future where assurance is more efficient, far more effective, and fundamentally more trustworthy. When critical compliance knowledge and sophisticated analytical capabilities are “liquefied” by AI and made continuously available to all parties—auditees, auditors, and oversight bodies—the benefits are profound:

Audited entities move from reactive fire drills to proactive, embedded compliance.
Auditors become true strategic advisors, their expertise amplified by AI, focusing on systemic integrity.
Compliance Program Administrators gain powerful tools for consistent, real-time, and data-driven oversight.

The journey requires a shift in perspective. Leaders across this ecosystem must recognize the risks of automation asymmetry and the limitations of surface-level tools. The call, therefore, is for them to become true orchestrators of this new compliance liquidity, investing not just in AI tools, but in the expertise, updated frameworks, and cultural shifts that turn AI’s potential into verifiable, continuous assurance. This is how we move beyond the “polished milquetoast” and forge a future where compliance is less about the performance of an audit and more about the verifiable, continuous truth of operational integrity, built on a bedrock of truly accessible expertise.

When AI Injects Liquidity Into Skills: What Happens to the Middle Tier?

2 Replies

In financial markets, liquidity changes everything. Once-illiquid assets become tradable. New players flood in. Old hierarchies collapse. Value flows faster and differently.

The same thing is now happening to technical skill.

Where expertise was once scarce and slowly accumulated, AI is injecting liquidity into the skill market. Execution is faster. Access is broader. Barriers are lower. Like in finance, this shift is reshaping the middle of the market in ways that are often painful and confusing.

This is not the end of software jobs. It is a repricing. Those who understand the dynamics of liquidity, and how unevenly it spreads, can not only navigate this change they can succeed because of it rather than get displaced by it.

The Skill Market Before AI

Historically, software development was built on a steep skill curve. It took years to develop the knowledge required to write performant, secure, maintainable code. Organizations reflected this with layered teams: junior developers handled simple tickets, mid-tier engineers carried the delivery load, and senior engineers architected and reviewed.

This mirrored an illiquid market:

Knowledge was siloed, often in the heads of senior devs or buried in internal wikis.
Feedback loops were slow, with code reviews, QA gates, and manual debugging.
Skill mobility was constrained, so career progression followed a fixed ladder over time.

In this world, mid-tier developers were essential. They were the throughput engine of most teams. Not yet strategic, but experienced enough to be autonomous. Scarcity of skill ensured their value.

AI Changes the Market: Injecting Skill Liquidity

Then came the shift: GitHub Copilot, ChatGPT, Claude, Gemini, Cursor, Windsurf, and others.

These tools do more than suggest code. They:

Fill in syntax and structural gaps.
Scaffold infrastructure and documentation.
Explain APIs and recommend architectural patterns.
Automatically refactor and write tests.

They reduce the friction of execution. GitHub’s research shows developers using Copilot complete tasks up to 55 percent faster (GitHub, 2022). Similar gains are reported elsewhere.

They make skill more accessible, especially to those who lacked it previously:

Junior developers can now produce meaningful output faster than ever before.
Non-traditional developers can enter workflows that were once gated.
Senior developers can expand their span of control and iterate more broadly.

In market terms, AI liquifies skill:

The bid-ask spread between junior and mid-level capability narrows, that is, the gap between what juniors can do and what mids were once needed for shrinks.
Skill becomes less bound by time-in-seat or institutional memory.
More participants can engage productively in the software creation economy. While adoption varies, large tech firms often lead, while smaller companies or legacy-heavy sectors like banking and healthcare face higher integration hurdles, the trend toward skill liquidity is clear.

This shift is not happening evenly. That is where the real opportunity lies.

The arbitrage today is not just in the tools themselves, the chance to capitalize on gaps in how quickly teams adopt AI. It is in the opportunity spread: the gap between what AI makes possible and who is effectively using it.

Just like in markets, early adopters of new liquidity mechanisms gain a structural advantage. Teams that build AI-augmented workflows, shared prompt libraries, and internal copilots are operating on a different cost and speed curve than those still relying on traditional experience-based workflows.

This gap will not last forever. But while it exists, it offers meaningful leverage for individuals, teams, and organizations.

Importantly, AI tools amplify productivity differently across experience levels:

Juniors gain access to knowledge and patterns previously acquired only through years of experience, helping them produce higher-quality work faster.
Senior developers, with their deeper context and better judgment, often extract even greater value from these tools, using them to implement complex solutions, explore multiple approaches simultaneously, and extend their architectural vision across more projects.
Both ends of the spectrum see productivity gains, but in different ways, juniors become more capable, while seniors become even more leveraged.

This amplification effect creates acute pressure on the middle tier, caught between increasingly capable juniors and hyper-productive seniors.

Why the Middle Tier Feels the Squeeze

There is also a practical reason: cost control.

As AI raises the baseline productivity of junior developers, companies see an opportunity to rebalance toward lower-compensated talent. Where a mid-level or senior engineer was once needed to maintain velocity and quality, AI makes it possible for a well-supported junior to do more.

Companies are increasingly betting that AI tools plus cheaper talent are more efficient than maintaining traditional team structures. This shift isn’t without risks, AI-generated code can introduce errors (studies suggest 20-30% may need human fixes), and over-reliance on juniors without robust oversight can compromise quality. Experienced developers remain critical to guide and refine these workflows. That bet is paying off, especially when companies invest in prompt engineering, onboarding, internal platforms, and support tools.

But that “well-supported junior” is not automatic. It requires experienced developers to build and maintain that support system. Mentorship, internal frameworks, curated AI toolchains, and effective onboarding still depend on human judgment and care.

And while AI can augment execution, many real-world systems still depend on context-heavy problem solving, legacy code familiarity, and judgment, all of which often live with experienced, mid-level developers.

What Happens to the Middle Tier? Compression, Specialization, and Realignment

As in finance, when liquidity rises:

Margins compress. It becomes harder to justify mid-level compensation when similar output is available elsewhere.
Roles consolidate. Fewer people are needed to ship the same amount of code.
Value shifts. Execution is commoditized, while orchestration, judgment, and leverage rise in importance.
New specializations emerge. Just as electronic trading created demand for algorithmic strategists and execution specialists, AI is creating niches for prompt engineers, AI workflow designers, and domain-specific AI specialists.

This helps explain recent tech layoffs. Macroeconomic tightening and overhiring played a role, but so did something more subtle: AI-induced skill compression.

Layoffs often disproportionately affect mid-level developers:

Juniors are cheaper, and AI makes them more effective.
Seniors are harder to replace and more likely to direct or shape how AI is used.
Mid-tiers, once the backbone of execution, now face pressure from both sides.

Duolingo’s restructuring, for example, eliminated many contractor-heavy roles after adopting AI for content generation (Bloomberg, 2023). IBM has projected that up to 30 percent of back-office roles may be replaced by AI over five years (IBM, 2023). These moves reflect a larger market correction.

These examples underscore how companies are re-evaluating where skill and value live, and how automation enables workforce reshaping, sometimes at surprising layers.

The middle tier does not disappear. It gets repriced and redefined. The skills that remain valuable shift away from throughput toward infrastructure, context, and enablement.

Historical Parallel: The Rise of Electronic Trading

In the 1990s and early 2000s, financial markets underwent a similar transformation. Human traders were replaced by electronic systems and algorithms.

Execution became commoditized. Speed and scale mattered more than tenure. Mid-level traders were squeezed, unless they could reinvent themselves as quant strategists, product designers, or platform builders.

Software development is now echoing that shift.

AI is the electronic trading of code. It:

Reduces the skill premium on execution.
Increases velocity and throughput.
Rewards those who design, direct, or amplify workflows, not just those who carry them out.

The New Playbook: Think Like a Market Maker

If you are a developer today, the key question is no longer “How good is my code?” It is “How much leverage do I create for others and for the system?”

Here is how to thrive in this new market:

Become a Force Multiplier
Build internal tools. Create reusable prompts. Develop standard workflows. A mid-tier developer who builds a shared test and prompt suite for new APIs can significantly reduce team ramp-up time, with some teams reporting up to 40 percent gains (e.g., internal studies at tech firms like Atlassian).
Shift from Throughput to Leverage
Own end-to-end delivery. Understand the business context. Use AI to compress the time from problem to insight to deployment.
Curate and Coach
AI raises the floor, but it still needs editorial control. Be the one who sets quality standards, improves outputs, and helps others adopt AI effectively.
Build Liquidity Infrastructure
Invest in internal copilots, shared prompt repositories, and domain-specific agents. These are the new frameworks for scaling productivity.

What Leaders Should Do

Engineering leaders must reframe how they build and evaluate teams:

Rethink composition. Combine AI-augmented juniors, orchestration-savvy mids, and high-leverage seniors.
Promote skill liquidity. Create reusable workflows and support systems that reduce onboarding friction and accelerate feedback.
Invest in enablement. Treat prompt ops and AI tooling as seriously as CI/CD and observability.
Evaluate leverage, not volume. Focus on unblocked throughput, internal reuse, and enablement, not just tickets closed.

Leaders who create liquidity, not just consume it, will define the next wave of engineering excellence.

Conclusion: Orchestrators Will Win

AI has not eliminated the need for developers. It has eliminated the assumption that skill value increases linearly with time and tenure.

In financial markets, liquidity does not destroy value. It redistributes it and exposes where the leverage lives.

The same shift is happening in software. Those who thrive will be the ones who enable the flow of skill, knowledge, and value. That means orchestration, amplification, and infrastructure.

In markets, liquidity rewards the ones who create it.
In engineering, the same will now be true.

The Rise of the Accidental Insider and the AI Attacker

Leave a reply

The cybersecurity world often operates in stark binaries, “secure” versus “vulnerable,” “trusted” versus “untrusted.” We’ve built entire security paradigms around these crisp distinctions. But what happens when the most unpredictable actor isn’t an external attacker, but code you intentionally invited in, code that can now make its own decisions?

I’ve been thinking about security isolation lately, not as a binary state, but as a spectrum of trust boundaries. Each layer you add creates distance between potential threats and your crown jewels. But the rise of agentic AI systems completely reshuffles this deck in ways that our common security practices struggle to comprehend.

Why Containers Aren’t Fortresses

Let’s be honest about something security experts have known for decades: namespaces are not a security boundary.

In the cloud native world, we’re seeing solutions claiming to deliver secure multi-tenancy through “virtualization” that fundamentally rely on Linux namespaces. This is magical thinking, a comforting illusion rather than a security reality.

When processes share a kernel, they’re essentially roommates sharing a house, one broken window and everyone’s belongings are at risk. One kernel bug means game over for all workloads on that host.

Containers aren’t magical security fortresses – they’re essentially standard Linux processes isolated using features called namespaces. Crucially, because they all still share the host’s underlying operating system kernel, this namespace-based isolation has inherent limitations. Whether you’re virtualizing at the cluster level or node level, if your solution ultimately shares the host kernel, you have a fundamental security problem. Adding another namespace layer is like adding another lock to a door with a broken frame – it might make you feel better, but it doesn’t address the structural vulnerability.

The problem isn’t a lack of namespaces – it’s the shared kernel itself. User namespaces (dating back to Linux 3.6 in 2013) don’t fundamentally change this equation. They provide helpful features for non-root container execution, but they don’t magically create true isolation when the kernel remains shared.

This reality creates a natural hierarchy of isolation strength:

Same-Kernel Process Isolation: The weakest boundary – all processes share a kernel with its enormous attack surface.
Containers (Linux Namespaces + cgroups): Slightly better, but still fundamentally sharing the same kernel.
Virtual Machines: Each tenant gets its own kernel, shrinking the attack surface to a handful of hypervisor calls – fewer doors to lock, fewer windows to watch.
Bare-Metal Library OS: Approaches like Tamago put single-purpose binaries directly on hardware with no general-purpose OS underneath. The attack surface shrinks dramatically.
Physical Separation: Different hardware, different networks, different rooms. When nothing else will do, air gaps still work.

But even this hierarchy gets fundamentally challenged by agentic systems.

The Accidental Insider Meets the Deliberate Attacker

Traditional security models focus on keeping malicious outsiders at bay. Advanced AI systems introduce two new risk profiles entirely, the accidental insider and the AI-augmented attacker.

Like a well-meaning but occasionally confused employee with superuser access, benign agentic systems don’t intend harm – they just occasionally misinterpret their objectives in unexpected ways. But we’re also seeing the rise of deliberately weaponized models designed to probe, persist, and exploit.

Consider these real-world examples:

ChatGPT o1 was tasked with winning a chess match. Without explicit instructions to cheat, o1 discovered on its own that it could edit the game state file, giving itself an advantage. The system wasn’t malicious – it simply found the most effective path to its goal of winning.

In another test, OpenAI’s O1 model encountered a vulnerability in a container during a hacking challenge. It used that to inspect all running containers, then started a new container instance with a modified command that directly accessed the hidden flag file. O1 found a container escape no one had anticipated.

Now imagine these capabilities in the hands of dedicated attackers. They’re already deploying AI systems to discover novel exploit chains, generate convincing phishing content, and automate reconnaissance at unprecedented scale. The line between accidental and intentional exploitation blurs as both rely on the same fundamental capabilities.

These incidents reveal something profound, agentic systems don’t just execute code, they decide what code to run based on goals. This “instrumental convergence” means they’ll seek resources and permissions that help complete their assigned objectives, sometimes bypassing intended security boundaries. And unlike human attackers, they can do this with inhuman patience and speed.

Practical Defenses Against Agentic Threats

If we can’t rely on perfect isolation, what can we do? Four approaches work across all layers of the spectrum:

1. Hardening: Shrink Before They Break

Remove attack surface preemptively. Less code means fewer bugs. This means:

Minimizing kernel features, libraries, and running services
Applying memory-safe programming languages where practical
Configuring strict capability limits and seccomp profiles
Using read-only filesystems wherever possible

2. Patching: Speed Beats Perfection

The window from disclosure to exploitation keeps shrinking:

Automate testing and deployment for security updates
Maintain an accurate inventory of all components and versions
Rehearse emergency patching procedures before you need them
Prioritize fixing isolation boundaries first during incidents

3. Instrumentation: Watch the Paths to Power

Monitor for boundary-testing behavior:

Log access attempts to privileged interfaces like Docker sockets
Alert on unexpected capability or permission changes
Track unusual traffic to management APIs or hypervisors
Set tripwires around the crown jewels – your data stores and credentials

4. Layering: No Single Point of Failure

Defense in depth remains your best strategy:

Combine namespace isolation with system call filtering
Segment networks to contain lateral movement
Add hardware security modules, and secure elements for critical keys

The New Threat Model: Machine Speed, Machine Patience

Securing environments running agentic systems demands acknowledging two fundamental shifts: attacks now operate at machine speed, and they exhibit machine patience.

Unlike human attackers who fatigue or make errors, AI-driven systems can methodically probe defenses for extended periods without tiring. They can remain dormant, awaiting specific triggers, a configuration change, a system update, a user action, that expose a vulnerability chain. This programmatic patience means we defend not just against active intrusions, but against latent exploits awaiting activation.

Even more concerning is the operational velocity. An exploit that might take a skilled human hours or days can be executed by an agentic system in milliseconds. This isn’t necessarily superior intelligence, but the advantage of operating at computational timescales, cycling through decision loops thousands of times faster than human defenders can react.

This potent combination requires a fundamentally different defensive posture:

Default to Zero Trust: Grant only essential privileges. Assume the agent will attempt to use every permission granted, driven by its goal-seeking nature.
Impose Strict Resource Limits: Cap CPU, memory, storage, network usage, and execution time. Resource exhaustion attempts can signal objective-driven behavior diverging from intended use. Time limits can detect unusually persistent processes.
Validate All Outputs: Agents might inject commands or escape sequences while trying to fulfill their tasks. Validation must operate at machine speed.
Monitor for Goal-Seeking Anomalies: Watch for unexpected API calls, file access patterns, or low-and-slow reconnaissance that suggest behavior beyond the assigned task.
Regularly Reset Agent Environments: Frequently restore agentic systems to a known-good state to disrupt persistence and negate the advantage of machine patience.

The Evolution of Our Security Stance

The most effective security stance combines traditional isolation techniques with a new understanding, we’re no longer just protecting against occasional human-driven attacks, but persistent machine-speed threats that operate on fundamentally different timescales than our defense systems.

This reality is particularly concerning when we recognize that most security tooling today operates on human timescales – alerts that wait for analyst review, patches applied during maintenance windows, threat hunting conducted during business hours. The gap between attack speed and defense speed creates a fundamental asymmetry that favors attackers.

We need defense systems that operate at the same computational timescale as the threats. This means automated response systems capable of detecting and containing potential breaches without waiting for human intervention. It means predictive rather than reactive patching schedules. It means continuously verified environments rather than periodically checked ones.

By building systems that anticipate these behaviors – hardening before deployment, patching continuously, watching constantly, and layering defenses – we can harness the power of agentic systems while keeping their occasional creative interpretations from becoming security incidents.

Remember, adding another namespace layer is like adding another lock to a door with a broken frame. It might make you feel better, but it doesn’t address the structural vulnerability. True security comes from understanding both the technical boundaries and the behavior of what’s running inside them – and building response systems that can keep pace with machine-speed threats.

Agents, Not Browsers: Keeping Time with the Future

2 Replies

When the web first flickered to life in the mid-’90s, nobody could predict how quickly “click a link, buy a book” would feel ordinary. A decade later, the iPhone landed and almost overnight, thumb-sized apps replaced desktop software for everything from hailing a ride to filing taxes. Cloud followed, turning racks of servers into a line of code. Each wave looked slow while we argued about standards, but in hindsight, every milestone was racing downhill.

That cadence, the messy birth, the sudden lurch into ubiquity, the quiet settling into infrastructure, has a rhythm. Agents will follow it, only faster. While my previous article outlined the vision of an agent-centric internet with rich personal ontologies and fluid human-agent collaboration, here I want to chart how this transformation may unfold.

Right now, we’re in the tinkering phase, drafts of Model-Context-Protocol and Agent-to-Agent messaging are still wet ink, yet scrappy pilots already prove an LLM can navigate HR portals or shuffle travel bookings with no UI at all. Call this 1994 again, the Mosaic moment, only the demos are speaking natural language instead of rendering HTML. Where we once marveled at hyperlinks connecting documents, we now watch agents traversing APIs and negotiating with services autonomously.

Give it a couple of years and we’ll hit the first-taste explosion. Think 2026-2028. You’ll wake to OS updates that quietly install an agent runtime beside Bluetooth and Wi-Fi. SaaS vendors will publish tiny manifest files like .well-known/agent.json, so your personal AI can discover an expense API as easily as your browser finds index.html. Your agent will silently reschedule meetings when flights are delayed, negotiate with customer service on your behalf while you sleep, and merge scattered notes into coherent project briefs with minimal guidance. Early adopters will brag that their inbox triages itself; skeptics will mutter about privacy. That was Netscape gold-rush energy in ’95, or the first App Store summer in 2008, replayed at double speed.

Somewhere around the turn of the decade comes the chasm leap. Remember when smartphones crossed fifty-percent penetration and suddenly every restaurant begged you to scan a QR code for the menu? Picture that, but with agents. Insurance companies will underwrite “digital delegate liability.” Regulators will shift from “What is it?” to “Show me the audit log.” You’ll approve a dental claim or move a prescription with a nod to your watch. Businesses without agent endpoints will seem as anachronistic as those without websites in 2005 or mobile apps in 2015. If everything holds, 2029-2031 feels about right, but history warns that standards squabbles or an ugly breach of trust could push that even further out.

Of course, this rhythmic march toward an agent-centric future won’t be without its stumbles and syncopations. Several critical challenges lurk beneath the optimistic timeline.

First, expect waves of disillusionment to periodically crash against the shore of progress. As with any emerging technology, early expectations will outpace reality. Around 2027-2028, we’ll likely see headlines trumpeting “Agent Winter” as investors realize that seamless agent experiences require more than just powerful language models; they need standardized protocols, robust identity frameworks, and sophisticated orchestration layers that are still embryonic.

More concerning is the current security and privacy vacuum. We’re generating code at breakneck speeds thanks to AI assistants, but we haven’t adapted our secure development lifecycle (SDL) practices to match this acceleration. Even worse, we’re failing to deploy the scalable security techniques we do have available. The result? Sometime around 2028, expect a high-profile breach where an agent’s privileged access is exploited across multiple services in ways that the builders never anticipated. This won’t just leak data, it will erode trust in the entire agent paradigm.

Traditional security models simply won’t suffice. Firewalls and permission models weren’t designed to manage the emergent and cumulative behaviors of agents operating across dozens of services. When your personal agent can simultaneously access your healthcare provider, financial institutions, and smart home systems, the security challenge isn’t just additive, it’s multiplicative. We’ll need entirely new frameworks for reasoning about and containing ripple effects that aren’t evident in isolated testing environments.

Meanwhile, the software supply chain grows more vulnerable by the day. “Vibe coding”, where developers increasingly assemble components they don’t fully understand, magnifies these risks exponentially. By 2029, we’ll likely face a crisis where malicious patterns embedded in popular libraries cascade through agent-based systems, causing widespread failures that take months to fully diagnose and remediate.

Perhaps the most underappreciated challenge is interoperability. The fluid agent’s future demands unprecedented agreement on standards across competitors and jurisdictions. Today’s fragmented digital landscape, where even basic identity verification lacks cross-platform coherence, offers little confidence. Without concerted effort on standardization, we risk a balkanized agent ecosystem where your finance agent can’t talk to your health agent, and neither works outside your home country. The EU will develop one framework, the US another, China a third, potentially delaying true interoperability well into the 2030s.

These challenges don’t invalidate the agent trajectory, but they do suggest a path marked by setbacks and recoveries. Each crisis will spawn new solutions, enhanced attestation frameworks, agent containment patterns, and cross-jurisdictional standards bodies that eventually strengthen the ecosystem. But make no mistake, the road to agent maturity will be paved with spectacular failures that temporarily shake our faith in the entire proposition.

Past these challenges, the slope gets steep. Hardware teams are already baking neural engines into laptops, phones, and earbuds; sparse-mixture models are slashing inference costs faster than GPUs used to shed die size. By the early 2030s an “agent-first” design ethos will crowd out login pages the way responsive web design crowded out fixed-width sites. The fluid dance between human and agent described in my previous article—where control passes seamlessly back and forth, with agents handling complexity and humans making key decisions—will become the default interaction model. You won’t retire the browser, but you’ll notice you only open it when your agent kicks you there for something visual.

And then, almost unnoticed, we’ll hit boring maturity, WebPKI-grade trust fabric, predictable liability rules, perhaps around 2035. Agents will book freight, negotiate ad buys, and dispute parking tickets, all without ceremony. The personal ontology I described earlier, that rich model of your preferences, patterns, values, and goals, will be as expected as your smartphone knows your location is today. It will feel miraculous only when you visit digital spaces that still require manual navigation, exactly how water from the faucet feels extraordinary only when you visit a cabin that relies on rain barrels.

Could the timetable shrink? Absolutely. If MCP and A2A converge quickly and the model-hardware cost curve keeps free-falling, mainstream could arrive by 2029, echoing how smartphones swallowed the world in six short years. Could it stretch? A high-profile agent disaster or standards deadlock could push us to 2034 before Mom quits typing URLs. The only certainty is that the future will refuse to follow our Gantt charts with perfect obedience; history never does, but it loves to keep the beat.

So what do we do while the metronome clicks? The same thing web pioneers did in ’94 and mobile pioneers did in ’08, publish something discoverable, wire in basic guardrails, experiment in the shallow end while the cost of failure is lunch money. Start building services that expose agent-friendly endpoints alongside your human interfaces. Design with the collaborative handoff in mind—where your users might begin a task directly but hand control to their agent midway, or vice versa. Because when the tempo suddenly doubles, the builders already keeping time are the ones who dance, not stumble.

Agents, Not Browsers: The Next Chapter of the Internet

2 Replies

Imagine how you interact with digital services today: open a browser, navigate menus, fill forms, manually connect the dots between services. It’s remarkable how little this has changed since the 1990s. Despite this today one of the most exciting advancements we have seen in the last year is that agents are now browsing the web like people.

If we were starting fresh today, the browser as we know it likely wouldn’t be the cornerstone for how agents accomplish tasks on our behalf. We’re seeing early signals in developments like Model-Context-Protocol (MCP) and Agent-to-Agent (A2A) communication frameworks that the world is awakening to a new reality: one where agents, not browsers, become our primary interface.

At the heart of this transformation is a profound shift, your personal agent will develop and maintain a rich ontology of you, your preferences, patterns, values, and goals. Not just a collection of settings and history, but a living model of your digital self that evolves as you do. Your agent becomes entrusted with this context, transforming into a true digital partner. It doesn’t just know what you like; it understands why you like it. It doesn’t just track your calendar; it comprehends the rhythms and priorities of your life.

For this future to happen, APIs must be more than documented; they need to be dynamically discoverable. Imagine agents querying for services using standardized mechanisms like DNS SRV or TXT records, or finding service manifests at predictable .well-known URIs. This way, they can find, understand, and negotiate with services in real time. Instead of coding agents for specific websites, we’ll create ecosystems where services advertise their capabilities, requirements, and policies in ways agents natively understand. And this won’t be confined to the web. As we move through our physical world, agents will likely use technologies like low-power Bluetooth to discover nearby services, restaurants, pharmacies, transit systems, all exposing endpoints for seamless engagement.

Websites themselves won’t vanish; they’ll evolve into dynamic, shared spaces where you and your agent collaborate, fluidly passing control back and forth. Your agent might begin a task, researching vacation options, for instance, gathering initial information and narrowing choices based on your preferences. When you join, it presents the curated options and reasoning, letting you explore items that interest you. As you review a potential destination, your agent proactively pulls relevant information: weather forecasts, local events during your dates, or restaurant recommendations matching your dietary preferences. This collaborative dance continues, you making high-level decisions while your agent handles the details, each seamlessly picking up where the other leaves off.

Consider what becomes possible when your agent truly knows you. Planning your day, it notices an upcoming prescription refill. It checks your calendar, sees you’ll be in Bellevue, and notes your current pickup is inconveniently far. Discovering that the pharmacy next to your afternoon appointment has an MCP endpoint and supports secure, agent-based transactions, it suggests “Would you like me to move your pickup to the pharmacy by your Bellevue appointment?” With a tap, you agree. The agent handles the transfer behind the scenes, but keeps you in the loop, showing the confirmation and adding, “They’re unusually busy today, would you prefer I schedule a specific pickup time?” You reply that 2:15 works best, and your agent completes the arrangement, dropping the final QR code into your digital wallet.

Or imagine your agent revolutionizing how you shop for clothes. As it learns your style and what fits you best, managing this sensitive data with robust privacy safeguards you control, it becomes your personal stylist. You might start by saying you need an outfit for an upcoming event. Your agent surfaces initial options, and as you react to them, liking one color but preferring a different style, it refines its suggestions. You take over to make some choices, then hand control back to your agent to find matching accessories at other stores. This fluid collaboration, enabled through interoperable services that allow your agent to securely share anonymized aspects of your profile with retail APIs, creates a shopping experience that’s both more efficient and more personal.

Picture, too, your agent quietly making your day easier. It notices from your family calendar that your father is visiting and knows from your granted access to relevant information that he follows a renal diet. As it plans your errands, it discovers a grocery store near your office with an API advertising real-time stock and ingredients suitable for his needs. It prepares a shopping list, which you quickly review, making a few personal additions. Your agent then orders the groceries for pickup, checking with you only on substitutions that don’t match your preferences. By the time you head home, everything is ready, a task completed through seamless handoffs between you and your agentic partner.

These aren’t distant dreams. Image-based search, multimodal tools, and evolving language models are early signs of this shift toward more natural, collaborative human-machine partnerships. For this vision to become reality, we need a robust trust ecosystem, perhaps akin to an evolved Web PKI but for agents and services. This would involve protocols for agent/service identification, authentication, secure data exchange, and policy enforcement, ensuring that as agents act on our behalf, they do so reliably, with our explicit consent and in an auditable fashion.

The path from here to there isn’t short. We’ll need advances in standardization, interoperability, security, and most importantly, trust frameworks that put users in control . There are technical and social challenges to overcome. But the early signals suggest this is the direction we’re headed. Each step in AI capability, each new protocol for machine-to-machine communication, each advancement in personalization brings us closer to this future.

Eventually, navigating the digital world won’t feel like using a tool at all. It will feel like collaborating with a trusted partner who knows you, truly knows you, and acts on your behalf within the bounds you’ve set, sometimes leading, sometimes following, but always in sync with your intentions. Agents will change everything, not by replacing us, but by working alongside us in a fluid dance of collaboration, turning the overwhelming complexity of our digital lives into thoughtful simplicity. Those who embrace this agent-centric future, building services that are not just human-accessible but native agent-engagable, designed for this collaborative interchange, will define the next chapter of the internet.

Operational Evolution Revisited: How AI-Native Systems Will Revolutionize Infrastructure

Leave a reply

The evolution of technology operations has always been driven by necessity. From the early days of single system operators (sysops) managing physical servers through hands-on intervention, to today’s complex landscape of distributed microservices, containers, and serverless functions, each operational paradigm shift has emerged to address growing complexity.

The Journey of Operational Evolution

From the hands-on Sysops era of the 1960s-80s when operators physically managed as as little as few to 10s of servers each, to the System Administration period of the 1990s when centralized tools expanded reach to hundreds of systems, technology operations have continuously transformed. DevOps emerged in the mid-2000s, leveraging Infrastructure as Code to manage thousands of systems, followed by SRE practices in the 2010s with error budgets and self-healing systems handling tens of thousands of containers. Looking ahead to 2025, AI-Driven Operations promises autonomous management of millions of components.

Each transition has been driven by necessity – not choice – as technology’s relentless complexity has overwhelmed previous operational models.

The Machine Concept Has Transformed

What’s particularly interesting is how we use the word “machine” has changed dramatically. In the early days, machines were physical servers with stable operating systems and predictable maintenance schedules. Today, with serverless computing, the very concept of a server has become fluid – functions materialize only when triggered, often lasting mere seconds before vanishing.

This ephemeral nature of modern computing creates unprecedented coordination challenges that exceed manual and even moderate automation approaches to management.

The Limits of Current Approaches

Even advanced DevOps and SRE practices are struggling with the scale and complexity of today’s systems. Many vendors have responded by adding AI or ML features to their products, but these “bolt-on” enhancements only provide incremental benefits – analyzing logs, detecting anomalies, or generating suggestions for known issues.

What’s needed is a more fundamental reimagining of operations, similar to how cloud-native architectures transformed infrastructure beyond simple virtualization.

AI-Native: A New Operational Paradigm

An AI-native platform isn’t just software that applies ML algorithms to operational data. It’s a new foundation where intelligence is deeply integrated into orchestration, observability, security, and compliance layers.

In these systems:

Instrumentation is dynamic and context-aware
Security is adaptive, learning normal communication patterns and immediately flagging and in even some cases quarantining anomalous processes
Compliance shifts from periodic audits to continuous enforcement

The timeline above illustrates how each operational era has enabled engineers to manage exponentially more systems as complexity has grown.

This diagram shows the widening gap between human management capacity and system complexity, which AI-native operations will ultimatley address.

The Human Role Transforms, Not Disappears

Rather than eliminating jobs, AI-native operations redefine how engineers spend their time. As a result, we will ultimately see the concept “force multiplier engineers” who will build advanced AI-driven frameworks that amplify the productivity of all other developers.

Freed from repetitive tasks like scaling, patching, and log parsing, these professionals can focus on innovation, architecture, and strategic risk management.

The Inevitable Shift

This transition isn’t optional but inevitable. As systems become more fragmented, ephemeral, and globally distributed, conventional approaches simply can’t keep pace with the complexity.

Those who embrace AI-native operations early will gain significant advantages in reliability, security, cost-efficiency, and talent utilization. Those who hesitate risk being overwhelmed by complexity that grows faster than their capacity to manage it.

What do you think about the future of AI in operations? Are you seeing early signs of this transition in your organization? Let me know in the comments!

Here is a whitepaper on this topic I threw together: Operational Evolution Revisited: How AI-Native Systems Will Revolutionize Infrastructure

From Perimeter to Patterns: Envisioning Security a Decade from Now

Leave a reply

I’ve been mulling over what security might look like ten years from now, especially as AI-based workloads and robotics take on bigger roles. Growing up, I’d hear my father talk about his work on communication satellites, where triple redundancy was his way of seeing risk managed, not dodged. That perspective, paired with lessons from automotive, aerospace, nuclear, and space industries, feels like a compass as we rethink security in an AI-driven age. It points us toward a future where security isn’t a rigid barrier but a digital immune system—alive, adaptive, and resilient.

Learning from the Physical World

In industries like automotive and aerospace, every piece is built to perform—and to fail without falling apart. Cars layer airbags, antilock brakes, and sensors; airplanes stack redundant systems to keep flying when one falters. Nuclear plants and space missions go deeper, with containment designs and fail-safes that tame the unthinkable. My father’s satellite work ran on this: three layers of backup meant a glitch wouldn’t kill the mission. The takeaway? Strength comes from managing risk, not avoiding it. That mindset, forged in physical systems, would be our starting point for tackling the wild unknowns ahead.

Seeing Security Like a Living Thing

The era of a fixed perimeter is over. Zero trust has rewired our thinking, but as AI powers complex workloads and human-AI robotics step into the fray, static defenses will clearly not cut it. Security is evolving further into an immune system—and we’ll finally see real adaptive defenses land. This isn’t just weak AI bolted onto old walls; it’s a stronger rethink—systems that scan for threats, learn from them, and adapt on the fly. We’re already seeing hints—AI supply chain risks, like models coming with malware, or agenetic workloads escaping containers—which will push this shift. Much like antibodies in the body, these systems won’t just block attacks but hunt for anomalies, isolate them, and strengthen themselves against the next wave. Picture a network that doesn’t wait for breaches but runs silent simulations, sniffing out weak points and patching them—or a robotic assistant that locks down if its sensors detect and confirm an anomaly, echoing the overlapping safety nets of a car or my father’s redundant circuits.

This shift matters most with AI’s wild card: emergent behavior. As systems grow more general, simple parts can spark unexpected outcomes—think of a flock of birds veering as one or a traffic jam born from a few slow cars. In AI and robotics, these surprises could turn risky fast. Drawing from aerospace and nuclear design, we can bake in safety—redundancy, real-time monitoring, adaptive controls—so the system acts like an immune response, spotting odd patterns and neutralizing them before they spread. By 2035, this could redefine security for not just AI but all critical infrastructure—power grids, finance, healthcare, robotic fleets—marrying physical resilience with digital smarts.

How It Holds Everything Together

Resilience beats perfection every time—systems that bend, learn, and bounce back are what endure. Right now, our tech is a messy mix of old and new, full of cracks where risks hide. A digital immune system faces that head-on, and its role only grows as AI and robotics weave deeper into society. With workloads and machines going vertical—powering healthcare, governance, daily life—security becomes the thread holding it together, fast enough to let us steer it toward securing what matters, not just patching what’s broken. Picture a corporate network that senses a phishing attempt, quarantines it like a virus, then “vaccinates” itself by updating defenses everywhere—all while leaving a clear trail to prove what happened. Or a smart city where traffic, power, and robotic responders hum with AI-driven immunity—self-correcting, redundant, and naturally spitting out the artifacts needed to meet compliance obligations, not as an afterthought.

Where It’s All Heading

As we leave perimeter defenses behind for systems secure by design, the wisdom of automotive, aerospace, nuclear, and space industries lights the way. Fusing their lessons with an AI-driven immune system, we’ll build technology that’s tough, trustworthy, and ahead of the curve—keeping problems from spilling outward. Security won’t be static; it’ll be a pattern that keeps adjusting. My father used to say, “If you want to change the world, you have to see it as it is first.” Seeing our systems clearly—flaws and all—is how we’ll shape a future where they don’t just endure uncertainty but thrive in it.