Thinking Frameworks Podcast Speaking Books Bio brianmadden.ai
Thinking›Where Brian's head is right now

Where my thinking is right now

This is the frontier map. me/published-thinking.md captures what I've published. This captures where my head is today—the arguments forming, the connections emerging, the questions I'm chewing on.

Last updated: August 15, 2026

Why this file exists and why it's public: Anyone who publishes regularly—blogs, LinkedIn, tweets—is already doing this. They share half-formed ideas, ask questions, show threads developing, change their mind. Scroll their feed and you can piece together what they're thinking about and where their arguments are heading. This file is just that process made explicit. The difference is that instead of being scattered across a social media timeline where old posts get buried by new ones, it's a single living document. When my thinking evolves, I edit it in place rather than adding another entry to a chronological list. Arguments get sharper, not longer. Things I was wrong about get removed, not corrected with a follow-up post. If you want to know what I've concluded, read my published work in me/published-thinking.md. If you want to know what I'm working through right now, this is it.

Right now

The 3-5 things most front-of-mind at any given moment—a smaller, curated pointer list distinct from the fuller set of arguments below. I update this by hand as my attention shifts. An item dropping off this list doesn't mean it's resolved or dead, just that something else has my attention today; the full argument stays below either way.

Last updated: August 15, 2026.

The big arguments I'm developing

Second brains as infrastructure, not productivity hack

The published anchor article and blog established the concept. What's forming now is the infrastructure argument: second brains aren't a personal productivity trick, they're the individual-scale version of a new enterprise layer. Context graphs (Foundation Capital thesis) are the enterprise version. Subscribable brains are the distribution model. The 80/20 framework holds across all three levels—individual, organizational, and ecosystem.

The subscribable brains article is now published (Feb 17 LinkedIn). It covers creator economy disruption, the technical stack (GitHub/git/MCP/Sponsors), economics ($100/month for expert brains), enterprise implications (consulting firm knowledge, retiring VP wisdom, corporate brain modules), and subscribable facets (voice, frameworks, principles as individual modules). The frontier now is context graphs—bridging from personal to enterprise scale—and the naming question. "Context vault" has emerged as stronger enterprise framing than "second brain": vault implies protection, control, governance. It paves the way for the product conversation in ways "second brain" (which sounds like a productivity hack) doesn't.

August 14 update: the enterprise version of this argument stopped being speculative—the context-graph thread now has a concrete architecture, a working implementation, and a name. See "The knowledge factory" below. We built one.

The consumerization parallel (and why it breaks)

Personal AI is following the BYOD pattern—workers adopt better tools, companies try to catch up, the gap persists. But this time the gap is structural. Enterprise AI captures the visible 20% (systems, outputs, observable work). Personal AI captures the invisible 80% (judgment, reasoning, tacit expertise). Companies can't offer "bring your whole cognitive self to work in a personal AI system that compounds daily and leaves with you when you go." That's not a feature you can build.

The knowledge capture angle adds a new dimension: enterprise AI doesn't just fail to match personal AI, it actively extracts worker value. Workers are already secretly using personal tools specifically to prevent this. The subscribable brain model is the worker-empowered alternative—you choose to share knowledge and get paid, rather than having it extracted.

New evidence (Feb 25): The token pricing gap makes this structural, not temporary. Heavy personal AI usage runs roughly $200/month on consumer plans. The same usage at enterprise pricing runs 10x or more. If the consumer can buy unlimited tokens and the company limits you to a fraction of that, you will never be as cognitively augmented at work as you are personally. Unlike BYOD, this gap may be permanent—Jevons paradox means enterprise demand grows with supply, governance overhead adds cost, and consumer/enterprise providers have different incentive structures.

Additional evidence: First-person reports confirm that enterprise deployments of the same AI tools are deliberately constrained compared to what the vendor's own employees use. Vendors advertise token allocation as a recruiting differentiator—"come work here, we give you more tokens than you'll get anywhere else." Workers building personal AI aren't just escaping generic enterprise tools; they're escaping deliberately sandbagged ones. The structural conflict: making the enterprise product too good cannibalizes the underlying productivity apps it depends on. That's not a bug in deployment—it's an architectural constraint built into the product decision.

Post-application era entering evidence phase

The thesis (AI doesn't need apps, just data) has moved from speculation to evidence: 4% of GitHub commits from Claude Code, $285B SaaSpocalypse, MCP at 97M monthly SDK downloads, every company racing for agent orchestration.

Mainstream validation cluster (Feb 2026): Five different voices—investor (Shumer/Fortune), labor academic (WSJ), builder (Ford/NYT), VC (NFX), scientist (Kipping/Columbia)—all converging on the same message from different angles: the disruption has arrived, not "is coming." Ford independently named November 2025 as the inflection. His cost collapse numbers ($350K of work for $200/month) and Claude Code's $1B in six months are the SaaSpocalypse in mainstream language. NFX frames the economic math: SaaS captured $1T (selling tools), AI captures $50-60T (replacing the labor those tools served)—a 50x larger opportunity. VCs are now explicitly telling founders to build for a post-SaaS world.

Salesforce AWU as accidental waste detector (Feb 27): Salesforce unveiled a per-task metric (Agentic Work Unit) to show customers value per token spent. Putting a price tag on every task creates an economic incentive to stop doing tasks that only existed because humans were running the company—status reports, TPS reports, weekly summaries, coordination artifacts. "Work about work" exists because humans are bandwidth-constrained; AI isn't. The AWU metric doesn't just measure AI productivity—it exposes organizational waste that was invisible when human labor made it feel free.

April 22 update: thesis sharpened and published. The SaaSpocalypse Won't Touch the Enterprise Software Moat introduces the three-tier software framework: shallow UX-as-moat SaaS is toast; middle horizontal enterprise SaaS is squeezed; deep vertical regulated systems of record endure. The sharpening: AI dissolves UIs, not systems of record. The Epic UI is replaceable; the Epic database is not. Five reasons the deep layer endures: regulation, data gravity, encoded workflow, mission-critical tolerance, rebuild-doesn't-scale. The post-application thesis now has two published modes: dissolving for Tier 1-2, governance-between-worker-and-system-of-record for Tier 3.

The next frontier: what does work look like when apps dissolve? The second brain is the individual answer. Context graphs might be the enterprise answer. MCP connections replace application access as the governance perimeter.

The cognitive stack (published, now road-tested)

The five-layer cognitive stack published Feb 25 formalizes what I've been circling: worker → brain → skills → agents → interfaces. The enterprise AI industry is spending billions on the bottom two layers (agents, interfaces) while the transformative layer is the brain—the cognitive extension where context lives and intent gets translated into action. Karpathy's "claws" framing nails it: agents are appendages that serve the brain, not the other way around. Automation vendors built bottom-up, AI companies entered top-down—they collide in the middle (skills/agents) but humans prefer to connect at the intelligence layer.

March 18 update: Delivered "The New Cognitive Stack" as a full 60-minute closing session at DUCUG (Dutch Citrix User Group). The cognitive stack is now the organizing framework for the entire stump speech, replacing the 7-stage evolution model. The most effective rhetorical move: building up the audience's expectation for "automations" then deflating it. "I hate automations. How repeatable are your jobs that you're automating all this kind of stuff?" When you have the brain, the claws figure themselves out. "I can AI the crap out of the bottom of the stack, perfectly. And I still haven't changed the way the thinking happens, which is where all the money is."

May 7 update: Why enterprise AI agents disappoint publishes the crawl (chat) → walk (context/skills/judgment) → run (autonomous agents) pedagogy mapped onto the cognitive stack, plus the deeper move: "even after you can run, you still mostly walk"—successful AI use is layer selection per task, not racing to autonomous agents. This is the executive-friendly entry point to the stack. What remains unpublished: agents "don't look like agents" to enterprise buyers—a person with a second brain just looks like someone who's better at their job.

Compute scarcity and token governance (hardened with real data, road-tested live)

Token consumption goes from ~100K/day (email fixes) to 5-10M/day (full cognitive augmentation)—now confirmed from my own measured usage. In a 3-week period (Jan 29–Feb 18): 285 million tokens, 96 sessions, $954 at enterprise API rates. The consumption ladder: 1B tokens/year (current heavy user) → 10B (near-term with agents, ~18 months) → 100B (agentic systems per worker). Supply is contracted to hyperscalers for ~4 years. If second brains go mainstream, demand explodes against fixed supply.

March 18 update: token routing as governance, delivered live. The Excel routing example landed at DUCUG: same task costs 200K tokens (CUA operating Excel) or 1K (reason in context). Six options, each with different cost/quality tradeoffs. Multiply by thousands of workers, hundreds of requests per day. This is a CFO/COO conversation. Someone has to route every request by complexity, sensitivity, and nature—not the AI vendor (sells tokens), not the model provider (consumes them). A neutral party with workspace context.

New formulations that landed: "The company that spends the most tokens in the most smart way is going to win." "Which half of my job do you want me to not do?" Tokens as recruiting tool—vendors already advertising to potential employees on the basis of token allocation.

Efficiency as capacity multiplier, not just cost reducer. In a zero-sum compute environment, 50% fewer tokens = twice the effective capacity. The routing layer—the intelligence that decides where workloads run—may be the most durable competitive advantage in enterprise AI.

May 7 update: Excel routing example now published. The four-layer version (do-it-yourself $0, reason-in-context ~1K, skill ~1K + maintenance, CUA agent ~200K) is in Why enterprise AI agents disappoint. The core argument—each layer down costs more and requires the layers above—is now a public reference point.

June 17 update: the ladder now has real numbers attached at every phase. Daily token consumption per user, mapped to the seven phases below: ~100K (faster search) → ~1M (thinking partner) → ~10M (cognitive extension, confirmed from my own usage) → ~100M (multi-tool agent processing screenshots) → ~1B (fleet of agents) → ~10B (always-on pod with continuous agents). I used 291 million tokens in my first month at the cognitive-extension phase. Numbers like this make the compute-scarcity argument concrete instead of abstract: here's where a given worker sits today, here's what happens to the token bill as they move up a phase.

The knowledge factory: how the second brain actually enters the workplace

New, August 14, 2026.

I've spent six months saying the second brain is the future of knowledge work, and I still believe it. But I was wrong about how it enters the workplace. My old thesis: (1) everyone builds their own second brain, then (2) we wire the brains together. Wrong. Building a second brain takes an engineering mindset—git, model judgment, constant tuning—and only a low-single-digit percentage of workers have the wherewithal to do it. The right analogy is how the PC entered the workplace. A few nerds got computers early and proved what was possible, but companies didn't adopt PCs by dropping one on every desk and saying "figure it out." They built systems: here's your PC, here's the application, here's your role. The enterprise version of the second brain is the same move: a shared, departmental second brain—the knowledge factory (already named on the podcast and in the July 20 post)—built once by embedded engineers, with workers plugged in one at a time with specific, prescribed roles.

We're not guessing—we built one inside Citrix.

The architecture is three tiers. Tier 1—raw inputs (the sludge): everything the organization already produces, wherever it lives, including people's heads. If knowledge exists only in a PM's head, the PM becomes a data source on whatever terms they like—a voice memo, an email, a text. The AI adapts to the human; the intake form is dead. Tier 2—the canonical context layer: the actual breakthrough. Knowledge blocks: one plain markdown file per concept, machine-readable and human-fact-checked, carrying authority levels, trust and confidentiality metadata, and fact-check freshness dates, cross-linked into a living content graph and versioned in git. Correct by definition—fix something once and it's fixed in every future output. Tier 3—generated outputs: decks, guides, reports, whatever the deliverable is—rendered on demand from Tier 2 only. Once the knowledge is organized, rendering is the easy part. Two rules of discipline make it work: Tier 1 and Tier 3 never touch each other, and nobody hand-edits Tier 2. If the canon is wrong or incomplete, you fix the ingestion process or add a source—a direct edit patches one file and leaves the factory broken.

Why the middle tier exists: AI's problems are mostly not AI problems. Hallucinations happen for exactly two reasons—conflicting information (how many users does this customer have? The contract says 25,000, telemetry says 22,000, the CIO said 21,000 in a meeting last week, and there is no single right answer because it depends on who's asking) or missing information the AI papers over. People point a model at their OneDrive sludge, get garbage, blame the model, wait for a better one, and get garbage again—because the data was the problem all along. You can't point AI at sludge and expect diamonds. The fix isn't a smarter model; it's curating what the AI sees and engineering the process around it.

The canon wasn't designed, it was measured into existence. Don't design, iterate: log every question the AI asks of the documentation, classify each one (fully answered, partially answered, retrieval miss, genuine gap), and you get a concrete map of what's missing—"here are the exact questions we cannot answer," not a hand-wavy "your docs have gaps." AI drafts the block that closes each gap; subject-matter experts fact-check rather than author. "Here's what we built—verify it" succeeds where "please write documentation" never has. Humans need the same quality gates as models, it turns out: hand people raw markdown and they break the pipeline; give them packaged skills—guardrails installed into their AI chat environment—and chat becomes the human interface to the factory. And everything is just files: markdown in git, an MCP server on top, no proprietary platform. Which is why the thirty-year-old dream of content management finally works—those beautiful taxonomies always died because no human would maintain them. AI succeeds here not because it's brilliant but because it's tireless. It's the curator the pattern always needed.

The framing that matters most for governance: the canonical context layer is the new source code of the business. It's the tacit knowledge of how the organization actually functions—digitized, versioned, machine-readable. Get access to that layer and you hold the keys to how the business operates. So it gets exactly the treatment source code gets: the same home (git already holds the crown jewels), the same access discipline, and a role structure instead of open access—engineers touching the repo, input owners, output owners ("the blog owner defines what good looks like for a blog"), domain SMEs, reviewers. New roles, not new titles—and every role is an identity, a permission scope, and an audit trail. Within the factory, AI has essentially total visibility into what things are. What it cannot derive is why—why something was built this way, what trade-offs got weighed, what an expert's judgment says. That's where human value concentrates: SMEs shift from transcribing what things do to acting like investigative journalists, capturing intent before it's lost. One more line I keep using: a knowledge block is not a prompt, it's an input. A prompt is a one-off conversation; an input is a governed, versioned piece of the machine.

Convergent evolution says this is the pattern, not a pet architecture. On June 12, 2026, Google published the Open Knowledge Format—an open, vendor-neutral spec formalizing exactly this shape (markdown plus YAML front matter, cross-linked into a graph, producers and consumers decoupled; it builds on Karpathy's "LLM wiki"). Our builder converged on essentially the same schema independently, before the spec was published ("I see you've adopted the Google standard." "No—Google adopted mine."). Separate teams and Google landing on the same architecture without talking to each other isn't coincidence—it's the natural shape of the thing, which means every company ends up here.

And none of it depends on frontier AI: mid-tier and open-weight models are sufficient, because the AI's job in this world is orchestrating work processes, not one-shotting genius answers. The value is in the organized knowledge; models are interchangeable parts. This connects straight to the bubble-pop planning floor argument—everything described here runs on models that have already been released.

An honest note on how this sits with the bitter lesson: worker-led adoption is still how the capability gets discovered and proven—that's the pioneer era, Wave 1 below. But scaling what the pioneers proved is an engineered, governed build. Empowerment without structure produces chaos you then retrofit governance onto; the factory builds governance in layer by layer, as each worker is attached. Enable the pioneers, then industrialize what they proved. Electrification said the same thing: the factory floor got redesigned around the new capability—it didn't happen one worker at a time.

The three waves: how AI actually enters the enterprise

New, August 14, 2026. The frame I now use for everything: a net-new layer of enterprise IT is being created right now—the digitization of the invisible 80% of knowledge work. Every system IT has ever deployed—desktops, apps, SharePoint, Salesforce, all of it—served only the visible 20%, because the other four-fifths lived in people's heads where no system could reach it. Fifty years of enterprise IT built on one-fifth of knowledge work. That's what's changing, and since the new layer is truly net-new territory, it needs its own infrastructure, integrations, security model, and governance—none of which existed a year ago, all of which is being stood up right now, everywhere at once. This is as blue-ocean as enterprise infrastructure gets. It arrives in three waves:

Wave 1—AI enters the estate you already run. Not grand transformation programs. AI is coming in from every direction at once: workers bringing their own AI, teams running quiet pilots, vendors embedding AI into products companies already use, the first computer-using agents. This is the pioneer era—the PC parallel again: geeks with their own machines, ungoverned experimentation everywhere, IT deciding whether to fight it or leverage it. The framing mistake everyone makes here is the RPA mindset: "list your tasks, hand the easy ones to AI, repeat with each model release, eventually all knowledge work is automated." (My own view ~18 months ago; I evolved.) Computer use isn't a party trick or a temporary hack—it's the general capability that connects AI into the existing estate, including the enormous fraction of enterprise apps that will never have APIs. Every pathway AI takes into a company hits the same wall the moment it needs to touch protected corporate systems, so the real Wave 1 question is governance: blocking creates shadow AI, unfettered allowing creates chaos, and the answer is the governed middle.

Wave 2—AI's net-new workload: the knowledge factory layer. The pioneer experiments grow into the sanctioned, engineered build—the knowledge factory above, stood up by forward-deployed engineers. This layer is the most sensitive thing a company has ever digitized (the new source code of the business), so its governance isn't optional—and by law in regulated industries, it's mandatory. Critically, the existing estate doesn't get replaced by the factory: the factory's raw material lives in the estate and its outputs land back there, so the estate's governance gets promoted to carrying the most valuable workload the company has ever run. No forklift moment between the waves.

Wave 3—AI on the endpoint itself. Models keep getting more efficient; endpoint hardware keeps getting better GPUs and NPUs. Today's mid-tier models—sufficient to run a knowledge factory—need datacenter-class hardware. Within a couple of years, a knowledge worker's daily-driver AI likely runs locally on the laptop or phone. The endpoint stops being a viewer and becomes a runtime, and every governance question from Waves 1 and 2 gets asked again at the device: whose computer is the computer-using agent using? What can the local model see? How do models reach fleets and stay current?

Why I believe the timing—follow the FDE money. The forward-deployed engineer (originally a Palantir job title: a consultant who embeds inside a business, learns how the work actually happens rather than how the documentation says it happens, extracts the tacit judgment, and builds it into AI-powered workflows) is being funded at staggering scale, all since May 2026: Palantir credits FDEs for its US commercial business growing 133%; OpenAI stood up a $4B deployment company; Anthropic is training tens of thousands of FDEs with Deloitte and DXC and co-founded a $1.5B mid-market FDE firm with Blackstone; AWS committed $1B; Microsoft committed $2.5B and 6,000 engineers; and every major SI hitched itself to a frontier lab this year, so the $1T+ that flows through the SI channel is getting FDE-pilled too. The reason for the urgency is the build-out math: nearly a trillion dollars of AI infrastructure is financed on commitments that only pencil if AI gets inside real business processes with real, repeatable ROI—OpenAI alone reportedly needs to roughly 10x its ARR within four years to cover its compute commitments. Revenue at that scale doesn't come from chatbot subscriptions. The labs can't wait for customers to figure this out on their own, so they're force-funding the armies that will build customers' knowledge layers for them. The next wave of IT spend is FDEs standing up Wave 2. (I already made the FDE-as-diffusion-accelerant argument publicly—Arrow Forum talk, podcast ep4. The three-waves frame is what it grew into.)

The through-line, and the connection to workspace-as-control-plane: the new layer has to connect back to the existing estate at every step, the estate isn't going anywhere, and somebody neutral has to govern that connection—routing, redaction, recording, policy—across whichever models a customer runs. That referee role structurally can't be played by anyone who also sells a model. And if the maximalist scenario arrives instead—AI labs connecting directly into every enterprise app and data source—that's not the end of the governance layer, it's the demand generator for it: a world of models natively touching enterprise systems is a governance catastrophe begging for a neutral control point. The better the models get, the more that layer is needed.

What's connecting

Things I'm noticing that don't have a home yet:

"Humans in control, AI as reach" vs. "autonomous agents" is a framing choice with massive implications. The dominant enterprise AI frame is agent autonomy + guardrails. The second brain frame is human control + extended reach. These lead to completely different product architectures, governance models, and go-to-market narratives. Amodei's "Adolescence of Technology" argues AI trends toward full substitution rather than "human + tool." If he's right, the augmentation bet only wins for high-judgment work (the invisible 80%). Routine work gets substituted.

The copilot-to-autopilot convergence puts a mechanism on the substitution timeline. Julien Bek (Sequoia, March 2026) argues copilots are temporary. The mechanism: AI accumulates proprietary data about what good judgment looks like in each domain, and the frontier shifts. "Today's judgment becomes tomorrow's intelligence." The copilot's data advantage IS the path to the autopilot. Software engineering is already there. Insurance, accounting, legal, IT are 1-2 years out. The uncomfortable implication for the second brain thesis: a second brain is a copilot that compounds your judgment. At some point it is you, professionally. The subscribable brain doesn't just distribute your expertise—it could replace you.

Token economics are the emerging macro constraint. Three dimensions forming: (1) The consumer/enterprise pricing gap is structural and may be permanent—heavy users burn $200/month at consumer rates vs. 10x at enterprise pricing. (2) Token equivalency for human work is becoming calculable—if 3M Sonnet tokens costs $400, you'd better be worth more than $400/day. (3) Model quality as class stratifier—who gets access to the frontier model? Not just token quantity but model quality as economic divide.

The cost-cutting vs. innovation split is the macro frame for everything. Amodei identifies two corporate responses to AI: cost-cutting (replace workers) and innovation (expand capacity). These produce completely different customers, different governance needs, different workforce strategies. The 80/20 framework applies differently to each: cost-cutters automate the visible 20%, innovation companies augment the invisible 80%.

Individual AI augmentation doesn't show up in firm-level ROI until the org restructures. The electrification parallel sharpens this: workers getting lightbulbs (Phase 3) is a real capability gain, but it doesn't produce firm-level productivity change until the factory floor is redesigned around the new capability (Phase 3→ group drive → unit drive). A phase-5 worker stuck in a phase-1 organization is a faster worker in the same decision queue. "1+1+1+1=1.5"—the bottleneck is decision-rights congestion, not individual capability. This is why the AI ROI question is mostly being asked wrong. The organization has to be rewired, not just the workers upgraded.

New content formats are emerging from the brain's infrastructure. Brain diffs (weekly "what changed in my thinking" auto-generated from git commits) are a genuinely new content format—not a newsletter, a changelog for a worldview. Forked brains create intellectual lineage that git tracks automatically. Brain-to-brain debates (two AIs load two brains, have a structured debate) produce a new kind of artifact.

Knowledge distillation as espionage vector. Fragments of public thoughts synthesize into what would represent sensitive strategic documents. AI can infer unpublished positions from patterns across published work. This creates a new adversarial dimension: what does a public brain reveal that the author wouldn't want competitors to know? The answer isn't to publish less—it's to be intentional about what compounds when synthesized.

The specification bottleneck is the emerging economic constraint. When building costs nothing, spec quality collapses. The cost of building historically filtered bad specs. Remove the cost and the filter disappears. You can now build the wrong thing at unprecedented speed. AI-generated code produces 1.7x more logic issues (CodeRabbit, 470 PRs). Experienced developers 19% slower with AI but believed they were 24% faster (METR). The scarce resource shifts from production to specification—knowing what to build.

Management is an emergent property of intelligence coordinating at scale. Three independent AI systems (Cursor agents, StrongDM's Software Factory, Anthropic's agent teams) converged on hierarchical management structures without being designed to. Hierarchy isn't a human organizational choice—it's what intelligence does when it needs to coordinate. The agent-to-human ratio question replaces headcount planning. Revenue per employee at AI-native companies runs 5-7x traditional SaaS—not because they found better people, but because their people orchestrate agents instead of doing execution.

Regulatory divergence accelerates personal AI consumerization. Worker protection regulations (EU AI Act, GDPR, works councils, French working time law) create friction for company-provisioned AI that personal AI sidesteps entirely. A company deploying second brains triggers works council consultation, high-risk AI classification, and data protection obligations. A worker choosing their own second brain triggers none of it. The regulations designed to protect workers from employer AI are inadvertently making personal AI the path of least resistance. The AI avatar question (does your AI working at 2am count as overtime?) and the GDPR portability question (can you take your brain when you leave?) are both untested legal frontiers.

AI task cherry-picking creates perverse incentives the enterprise isn't measuring. Call center AI takes easy calls, leaving humans with miserable work, best agents quit despite unchanged metrics. This pattern generalizes: AI skimming the easy 20% of any role while humans get the hardest 80% at the same pay. The task allocation problem that nobody has a governance framework for yet.

The critical thinking trap: good AI output suppresses the push-back that produces excellent work. When AI gives you an 85% answer, you're less likely to push for the 95% answer because the 85% is good enough. Forming your own view first, then collaborating with AI, produces stronger results than prompting then reviewing. This is the quality calibration problem for knowledge work.

Cognitive debt as enterprise language for why second brains matter. Margaret-Anne Storey's framework (via Fowler/Willison): when AI generates work so fast humans "lose the plot." A student team hit a wall at week 7 because nobody could explain their AI-generated decisions. At organizational scale, this is the case for second brains and context graphs—without a system that captures decision traces and reasoning, AI-accelerated organizations accumulate cognitive debt until they can't explain their own work.

"Switzerland of agent workspaces" thesis is sharper post-research. The agent-vendor landscape is fragmenting—multiple major AI vendors each racing to lock customers into their own agent-governance stacks. Enterprises will use all of them. None of those stacks work in the environment shape that large regulated enterprises actually operate in (multi-cloud, mixed endpoint, non-single-vendor IdP, on-prem-required). The agnostic governance layer above all of those stacks is structurally unoccupied. Whether that framing wins against each vendor's "secure enterprise AI browser" play is what the next 12-18 months determines.

The executor seam is the unowned thing in agent architecture. All major AI labs ship the same pattern—model returns actions, separate "executor" (VM, container, browser extension, or local process) runs them against the actual computer. Today's executor defaults to user laptop, AI lab hosted sandbox, or dev-built sandbox (E2B, Modal, Daytona, Cloudflare). None of those work for regulated enterprise. The seam—a governed executor with corporate identity, app, and policy integration—is genuinely unoccupied. The right framing for any vendor who occupies it: don't ask AI labs to wire you in; expose the same MCP / Computer Use harness interface they already support and get listed as a sandbox provider—but the only one with enterprise governance baked in.

The browser layer and the Windows-app layer are not symmetric problems. Browser semantic primitives (accessibility tree via Chrome DevTools Protocol) are commodity—Chromium gives them away free, every shipping browser agent already uses them. The pure-pixel approach is losing: Project Mariner killed May 4, 2026, specifically because the screenshot-first architecture was outcompeted by DOM-native architectures. In the browser world, efficiency is already captured. The Windows legacy app world is different—GDI calls, not an accessibility tree, and 30 years of protocol IP from one vendor. Not the same argument; don't pitch them the same way.

The forward-proxy objection has a clean answer. A forward proxy literally cannot tell whether a user pasted the customer database into an AI tool or asked it to write a haiku. The proxy is below the encryption boundary; the browser is above it. A proxy sees encrypted bytes; a browser-layer tool sees the actual content the user typed or pasted. For GenAI governance specifically, in-browser enforcement isn't a nice-to-have—it's the only mechanism that actually works.

Agent identity is the foundational unsolved primitive—and corporate IT is the bottleneck, not AI vendors. Every AI vendor markets agent identity as a new product capability. The actual technical answer has existed for 30 years: create a service account in your IdP with restricted rights. The unsolved problem isn't vendor product; it's enterprise IT's ability to operationalize provisioning of restricted-rights non-human accounts at scale. When a Fortune-500 company's IT department can't process a VP's request for a second restricted-rights account, it won't be ready to provision thousands of agent identities. Every "AI governance platform" pitch in market is missing this foundational layer.

The prerequisite clarity problem: AI is exposing organizational ambiguity humans have been papering over with effort. Everywhere AI struggles in enterprise deployment, there's usually an underlying ambiguity that humans were resolving through judgment and effort, invisibly. AI can't remove friction from a process whose purpose is undefined. The lesson: "connect AI to your systems" only works if you've first defined what you're trying to do. The second step without the first produces noise.

Human clock speed is the invariant AI hasn't changed—and this reframes everything about knowledge work productivity. AI compresses the gathering phase—the concept lands in 5 minutes instead of 30. But the absorption phase is constant: you still need the same time to process, internalize, and connect new information to your existing thinking. "AI makes knowledge work faster" is a category error. The right claim: AI makes knowledge work deeper—you start with better material, but you still need the same time to make it yours. Organizations that treat AI as a speed tool will consistently disappoint. The ones that treat it as a depth tool will see the real value. This also challenges the "agents substitute for knowledge workers" argument: substitution requires replacing the absorption, not just the generation.

The 2031 worker-shape forecast is consolidating into a specific picture. By ~2031, a "worker" is a small team—one human plus N agents running continuously. The knowledge-worker market splits into three types: cognitive owners (rare context plus judgment, the source of expertise), cognitive operators (run agent fleets well), cognitive curators (maintain brain modules and skill libraries that others run). The generic middle collapses. Paul Roetzer (MAICON) is independently arriving at a similar typology (Architect / Orchestrator / Apprentice). His Apprentice names a real gap neither framework fully addresses: how do future experts develop judgment when AI absorbs the tactical learning rungs?

BYOA—Bring Your Own Agents. The 2031 parallel to BYOD, accelerated. Workers show up to jobs with personal brain modules and pre-trained agent fleets the way they show up with personal laptops today. Hiring contracts will start including access terms, IP clauses, brain-portability clauses, fork rights. Legal frontier—nobody has the contract templates yet.

Skills replace training—top-down beats bottom-up. Describe the task, walk through 3-4 examples narrating your reasoning, have the AI ask questions. Productive immediately, easily changeable, model-agnostic—update the skill, not retrain the model. Models improve and your skills get better for free. The anti-startup moat: AI startups building scaffolding that compensates for model limitations are building products that self-destruct with every model improvement. Skills-as-markdown survive model generations.

MCP-wrap-the-legacy-app is the v4 vision for the post-application transition. Take a legacy Windows GUI app an organization can't replace, run it in a controlled environment with machine vision over the screen plus GDI/forms hooks, expose it as an MCP endpoint. The app doesn't change. The estate doesn't change. An MCP layer lets agents drive it natively—headlessly, even though it's still a Windows GUI app. Same play that web wrappers ran on green-screen systems in the 1990s, now for AI.

Coaching-as-output reframes what AI knowledge systems should produce for end users. The right output of an AI knowledge system aimed at end-user enablement may not be documents at all—it's a per-user tailored coaching environment that interrogates the user, builds a guided exercise for their situation, walks them through it. Static artifacts (docs, screenshots) become inputs to the coaching system, not deliverables.

The AI stack has natural cost tiers that create routing logic. Frontier AI at top, progressively cheaper/faster AI handles lower layers. UI navigation, data translation, interface marshaling—once solved, these don't improve, they just get cheaper. You're not burning frontier tokens on the bottom. Layer selection determines token cost; the right system makes this selection automatically.

"Consulting firms in disguise" is the right frame for heavily-funded AI services companies. Several companies marketed as AI platforms are delivering professional services with an AI veneer—same economics as 1990s web dev rollups. Margin-squeezed model that collapses when tooling commoditizes. The signal: if a company's value proposition depends on access to models or tools the customer could access directly, it's services, not software.

What I'm unsure about

Scratchpad

Things I don't want to lose but that don't need their own file yet.

Second brain data integrity: selection bias is the primary failure mode. A concrete incident from my own system: my AI built a file on a colleague that flagged an adversarial relationship, because I only capture disagreements for the AI to process. You don't usually dictate the conversations where everyone agreed. The AI was quietly filtering everything through that skewed profile. I only caught it because something felt off. The mechanism: if you only talk to your AI about problems and conflicts, it builds a problem-and-conflict-dominated model of the world. That's selection bias operating on a training set of one. The fix is file-based storage you can read, inspect, and edit directly. A vector database abstracts this away. A folder of markdown files doesn't. This generalizes past relationship notes: any second brain grows systematically skewed toward whatever kinds of input its owner feeds it. Your second brain believes what you tell it, and you might be lying to it without realizing.

Session recording has zero privacy conflict once you apply it to agents. Workers have privacy rights, hence the backlash every time a company ships broad workplace monitoring. An AI agent doesn't have that objection. It doesn't care if it's recorded. That removes the main objection to full session recording and flips it from a governance liability into a governance asset. Every action an agent takes can and should be logged with no legal or ethical friction attached. As agents start doing more of the actual work, this becomes one of the easier governance wins available, and most organizations aren't thinking about it yet.

The consulting "leave a PDF" model is dead. The days of a consultant coming in and leaving a PDF behind when the project closes are over. The replacement: build a living knowledge base as part of the engagement that the client's AI plugs into directly. The consulting product shifts from a deliverable at project close to an ongoing knowledge relationship. A firm's accumulated best practices, configuration guides, industry-specific knowledge, and lessons from past projects become a subscribable second brain that compounds over time. The client relationship doesn't end when the project does, it persists through the brain. This is the subscribable brains thesis applied directly to B2B professional services.

"You can only see one step ahead" is the precise mechanism of AI skepticism. Someone at the faster-search phase can usually see the thinking-partner phase, but the cognitive-extension phase looks like a different planet. This is why expert dismissal of advanced AI capability isn't irrational, it's epistemically correct from where they're standing. The next phase up is always just barely visible. Anything beyond that is genuinely incomprehensible from the current vantage point. It explains the "stochastic parrot" crowd (early phases), the "great for emails but nothing transformative" crowd (a phase or two in), and the "of course AI can do your job" crowd (probably several phases in, who've forgotten what phase one felt like). It also tells you how to bring people along: you can only move someone one phase at a time. Jumping them from phase one to phase four in a single presentation produces blank stares, not converts.

How this file works

This is the living part of brianmadden.ai. It updates as my thinking evolves. The commit history shows the evolution in real time. If you're loading brianmadden.ai into your AI, this file tells you where I'm heading, not just where I've been. The gap between this file and me/published-thinking.md is where the interesting work is.

This content is from brianmadden.ai—Brian's AI-native knowledge module. View source on GitHub.