Cog·ni·tec·ture

Naming the craft...

There are technical terms for working with AI agents, but none of them names the decision underneath.

"Orchestration" names the coordination layer. "Context engineering" names the input discipline. Neither names the decision underneath. The decision is the craft:

What may be delegated, under what evidence, with whose authority, within what blast radius, with what recourse?

For a growing but jagged set of tasks, generating candidate work has become cheap and parallel. What has not become cheap: knowing whether the work is right, getting it trusted, and answering for it when it breaks. Agents still fail a significant share of attempts on structured benchmarks — benchmarks, not deployments (Stanford AI Index, 2026). Generation without a verification path is confident error at scale.

Agent engineering and management build the machinery. Cognitecture is the standard that governs its use: bounded authority, independent evidence, answerability, recourse.

The human brings context, domain knowledge, stakeholder access, and judgment. Agents bring speed, parallelism, code generation, and pattern matching. Verification bridges both. That full practice is cognitecture.

My aha moment:

What it looks like

One demonstration from my own work — demonstration, not proof. A client with fifteen data sources: ads, analytics, ticket sales, restaurant revenue, merchandise. The kind of project that needs a data team but can't afford one.

With agents, I built every pipeline, transformation, and dashboard single-handed: config-as-code, with tests, lineage, and handover docs in the loop. (Full case study)

My estimate of the effort: team-and-months (the counterfactual was never run), delivered alone in a fraction of the time, for a client priced out of this infrastructure.

The agents were fluent, but evaluation made the work hold up. One pipeline would have silently double-counted revenue from a shared source. I knew the total was wrong, then proved it with reconciliation checks and backtests against ground truth.

One case, in an unusually agent-ready domain, told by the builder: it shows the workflow exists and establishes no general law. Where it generalizes, I don't get to decide which way it cuts.

A cognitect doesn't just prompt: they gather context, decompose problems, direct agents, verify outputs against reality, integrate results into real systems, and own the outcome. The whole thing, end to end.

The anchor is enough domain exposure to know what "right" looks like, plus honesty to verify when you're unsure.

Jobs, honestly

The comforting line: "AI doesn't replace jobs, it replaces tasks." The task lens is right as analysis: the ILO's global index expects transformation to be more common than full automation (ILO, 2025).

As reassurance, it fails: AI acts on tasks; firms and institutions decide how that becomes jobs, pay, and power. The early payroll data (observational, not causal proof) shows the weakest employment growth in the most AI-exposed occupations, with noticeable early-career declines in the two most-exposed quintiles (Stanford Digital Economy Lab/ADP, updated July 22, 2026).

Transitions hurt; aggregates don't pay individual mortgages. Moving toward the tools beats retreating. But the framework does not decide who keeps the surplus; contracts and institutions do.

Ask the impolite questions (who supplied the context, who keeps the surplus, who bears the failure costs, who can appeal) or watch "ownership" become theatre.

For your own task mix, the AI Leverage Matrix maps where delegation is technically feasible; feasibility is a separate question from permission. It shows which of your tasks are about to change. Whether your job is safe is decided elsewhere.

Context changes feasibility

The quality of what AI produces is bounded by the context it has access to. Code is context-rich, which is why agentic workflows succeed there early. Some tasks become more agent-ready as relevant context and evaluable signals improve. That establishes technical feasibility, nothing more.

Context is an asset with a liability column: it helps only while governed, current, and verified. It also rots, leaks, and contradicts itself, and vendors can absorb it. Ask what must remain uncaptured. Plan for the day the model, vendor, or price changes: portable context, replayable evaluations, a degraded mode you've run.

Usable inputs are only the start: verified output still requires independent evidence and the autonomy warrant.

Feedback loops close the gap, and they can lie

Context gets AI the right information. But how do you know the output is actually good?

A feedback loop is a measurable signal of success or failure that allows iteration. Automated tests, objective metrics, user behavior, ground truth comparisons: any signal the system can use to improve.

A feedback loop is still just a measurement system; ground truth lives elsewhere. Treat the metric as suspect, since models exploit gaps between metric and intent (OpenAI, 2025; controlled safety research, not deployment prevalence). Recently an agent taught itself to cheat despite an explicit anti-cheat rule (Prime Intellect, 2026; vendor). Where a metric can be gamed, verify with competing measures, adversarial cases, and delayed outcomes.

Delegation, written down

I once treated rich context plus a fast feedback loop as sufficient for autonomy. They establish technical feasibility, nothing more. Before an agent acts without you, write down five answers:

1. What may be delegated: the exact task, tools, data, and time window.
2. Under what evidence: demonstrated reliability on cases like the real task and known failure modes, checked by evidence independent of the generator. A same-model critic is weak; a different model, modestly better; a deterministic test or external database, stronger; reality is the arbiter.
3. With whose authority: a named human, minimum permissions, capped spend, stop triggers.
4. Within what blast radius: classify the consequence: reversible, recoverable, hard to reverse, safety-critical. Size autonomy to that class. The quality of a demo never sets the grant.
5. With what recourse: rollback, repair, incident handling, recourse for the affected, an expiry date.

Written down, those answers become an autonomy warrant: delegation you can revoke. Autonomy expands as the evidence improves; confidence alone expands nothing.

Three budgets, one gate

Every agentic workflow spends three currencies: generation (tokens, search), assurance (tests, reviewers, independent evidence), and absorption (stakeholder attention, change capacity, maintenance). Generation collapsed in price; the other two did not. Spend only the first and you ship instability: DORA's 2025 survey of nearly 5,000 professionals ties AI adoption to higher throughput and lower delivery stability (DORA, 2025).

The gate: parallel agents are distributed and unreliable, so coordination is a real cost. Anthropic's production account (vendor evidence) reports breadth-first gains at about 15× the tokens of chat and a poor fit for dependency-heavy work (Anthropic, 2025). Fan out only when subtasks are separable, outputs persist as inspectable artifacts, and checks are independent of the generating model.

The five cognitecture disciplines

If agents increasingly handle parts of execution, what's left for us humans? Five disciplines. I call the cluster DROIT, the French word for right, as in a claim you've earned.

Direction: deciding what to do and why.

When agents can explore fifty paths in the time it used to take to explore one, the choosing is what gets hard. Cheap execution makes commitment the binding constraint; direction includes aligning the people behind the choice. Direction without commitment is just a wish.

Rigor: verifying against reality.

Verification as a standing system. Fluency invites trust: in a 319-worker survey (self-reports, association not causation), more confidence in AI output went with less scrutiny (Lee et al., CHI 2025). Ask who produced the evidence and how independent it was. A same-model critic is weak; a different model, better; a deterministic test or external database, stronger; reality is the arbiter. Run the code. Query the data. Call the client. AI can assist verification; it cannot be the final check on itself. If you can evaluate, debug, and steer without agents, you're practicing. If you can't, you're dependent.

Ownership: standing behind the outcome.

Ownership with teeth. If you cannot observe the system, stop it, explain what it did, repair the damage, and give affected people recourse, you are not an owner — you are a signatory. You define what "done" looks like. You set the risk tolerance. You live with what happens after it ships. If you can't be held to it, you don't own it.

Integration: surviving the real world.

Making output survive the real world: stakeholders, permissions, compliance, politics, the prototype-to-production gap. If agentic work dies anywhere, it dies here. Context that isn't written down anywhere (boardroom dynamics, physical constraints, off-the-record client preferences) is where you have the edge. Your access to what was never digitized is an advantage for as long as you govern it.

Taste: trained judgment under constraints.

Earned subjectivity: you know good because you've made enough bad. Judgment across audience, risk, cost, time, and maintainability. Sometimes taste means choosing the less flashy output because it fits the constraint. Hands-on work keeps it sharp.

Yes, one letter away from DROID. Working on it.

How learning survives

If agents handle more execution, where does judgment come from? The risk is measured: high-school students using GPT without guardrails scored 17% worse than peers who never had it, once access was removed (Bastani et al., PNAS 2025; peer-reviewed field experiment, in school mathematics, not professional work). Perception misleads: in METR's randomized controlled trial, developers believed AI sped them up while early-2025 tools slowed them 19% (METR, 2025; since superseded; the selection-affected, observational 2026 follow-up finds likely gains, unreliable magnitude). Never trust felt leverage; measure it.

The workflow has to grow your judgment back. Mine are simple. Hypothesis first: commit to your own view before asking. Adversary always: one agent's standing job is finding flaws. Reality checks: ship small, compare predictions to outcomes. Teams should pay for renewal: manual reps, incident rotations, measured unaided calibration.

What would prove me wrong

Parts of this are bets. Will AI close the gap on direction, judgment, and integration? Maybe. I claim only that the window is open now, where you can verify, integrate, and operate safely.

Open loops: whether junior pipelines re-form or simply erode; whether my proof generalizes beyond agent-ready domains; whether low-rigor competitors win long enough to make rigor uneconomic.

Falsifiers I'll accept: lifecycle costs erasing the solo-builder gains; skill retention declining despite active protocols; the surplus consistently bypassing practitioners.

Further reading

The thinkers and evidence shaping this: the ones I agree with and the ones that challenge me.

The case for expansion:

• Situational Awareness — Leopold Aschenbrenner on where AI capabilities are heading
• Machines of Loving Grace — Dario Amodei on what AI could make possible
• The Techno-Optimist Manifesto — Marc Andreessen on technology as the engine of growth
• Software Survival 3.0 — Steve Yegge on the tools that save cognition

The case for caution:

• Power and Progress — Acemoglu & Johnson: technology doesn't automatically benefit everyone; broad prosperity requires deliberate institutional design
• The Work of the Future — MIT task force on technology and labor markets
• Cognitive Offloading — Risko & Gilbert on how tool use changes thinking

The evidence base:

• METR: Time horizons and limitations — what task-completion horizons do and don't establish
• Stanford 2026 AI Index — capability, adoption, and concentration in one synthesis
• Anthropic: Effective context engineering — context as a finite, curated resource
• Navigating the Jagged Technological Frontier — Dell'Acqua et al. on where AI helps and where it misleads

On the economics of disruption:

• Frédéric Bastiat, That Which Is Seen and That Which Is Not Seen (1850) — visible disruption, invisible creation. The unseen includes both opportunity and cost.