The cognitive stack

A five-layer model showing how intelligence organizes itself to coordinate complex work—from human intent down to mechanical execution. The enterprise AI industry is spending billions on the bottom two layers. The real transformation lives at the top.

Published: February 25, 2026 — Original post Extends: the delegation-not-automation thesis (formalizes the skills hierarchy into a named framework) Incorporates: Karpathy's "claws" concept for personal AI agents

The five layers

1. The worker: States intent and exercises judgment. "Get the pre-reads to all participants before March 6." The human decides what matters, what's urgent, and what the goal actually is.
2. The cognitive extension ("the brain"): The thing the worker actually talks to. Holds full context: who the participants are, preferred formats, what's sensitive, what happened last time. Plans the approach and sequences the work.
3. Skills: What the brain knows how to do. Coherent chunks of capability—process a transcript, draft an email, research a competitor, check a calendar. Each handles a meaningful piece of work with some autonomy.
4. Agentic sub-processes: How skills get executed. The agents that reach into systems, navigate interfaces, call APIs, coordinate with other agents. This is where the "agent" hype lives—and it's the second lowest-value layer.
5. Interfaces: The simplest, most mechanical layer. APIs, MCP, CUA, RPA, A2A, file interfaces, connectors, webhooks, scripts. Important plumbing, but interchangeable commodity infrastructure.

Vocabulary update: "harness" for the differentiating middle

2026-09-04. The industry has converged on a name for layers 3-4 together—the skills and the agentic sub-processes that turn a commodity model into something that actually does work: the harness. Mitchell Hashimoto's February 2026 post coined "harness engineering," and within months it's standard usage—Claude Code's own docs describe itself as "the agentic harness around Claude," Hugging Face's glossary formally distinguishes harness/model/scaffold, and there's now a Wikipedia page for the term. The evidence that it's not just vocabulary: SemiAnalysis ranks harness quality above benchmark score for real-world outcomes (a lower-scoring model with a better harness beating a higher-scoring one), DeepSeek open-sourced a 160k-star harness framework, and AlphaSignal states the thesis outright—"model plus harness is becoming one unit." Adopting "harness" as the name for layers 3-4 here rather than contesting the term: it's a better fit than "skills" alone, which named only layer 3 and left layer 4's execution machinery unnamed.

Two trajectories colliding

The bottom-up trajectory: RPA and automation vendors started at layer 5 and have been climbing upward for decades. Each time automations hit a ceiling requiring judgment, they moved up a layer. Exception handling, decision logic, now AI for ambiguous cases.

The top-down trajectory: AI companies entered from layer 1. LLMs that understand intent can reason about tasks and decompose work on their own. Workers describe what they want in natural language and the system figures out the steps.

These trajectories overlap in the middle (layers 3 and 4). Both modern AI companies and legacy automation vendors cover the skills and agentic layers. But humans prefer to connect at the intelligence layer—the delegation thesis holds that workers don't care about automations and agents. They want to tell their AI what they want and have it happen.

The automation fallacy: why workers delegate instead of building

Absorbed 2026-09-13 from the standalone "delegation, not automation" framework. The enterprise AI pitch has been the same for decades—give workers tools to automate their own workflows. RPA, low-code, citizen development, AI automation studios. Each generation makes the tools easier and the connectors more numerous, and the result is always the same: the ~1% of workers who think like programmers build amazing things, and the other 99% keep doing their jobs the way they always have. That's not a failure of the tools, it's a failure of the premise. Knowledge workers don't have stable, repeatable workflows worth the effort of designing and maintaining—their work is fluid, reactive, different every day—and even when something could be automated, they don't want to stop working to figure out how. They want to hand it off and move on, the way a manager delegates.

This is why the industry keeps building from the wrong end (below): automation studios ask the 99% to become the 1%. Today's AI automation studios are the equivalent of building BlackBerry apps—powerful, useful for the few who invest the time, and beside the point once the real interface arrives. The iPhone didn't ask users to learn a new programming paradigm; it hid decades of technical magic behind an interface intuitive enough for anyone. AI gets there the same way: workers won't build automations, they'll talk to their AI, and if it needs to fire up an agent, open a browser, call an API, or trigger a workflow, it does—in the background, invisibly. The automation studios being built today aren't wasted effort; they're the plumbing—connectors, access patterns, integrations—that the cognitive layer will use. But they're stepping stones, not the destination.

Claws, not brains

Karpathy coined "claws" for personal AI agents—a play on OpenClaw. The analogy is precise: claws are appendages. They grip, manipulate, and reach into systems. But they serve the brain, which decides what needs to happen, in what order, and with what judgment.

This maps directly onto the stack: layers 4-5 are claws, layer 2 is the brain, and layer 3 (skills) is the interface between them. The worker sits above, exercising judgment. The claws execute below. The brain orchestrates.

Why the industry is building from the wrong end

Agent marketplaces, orchestration engines, workflow designers, automation studios—all investing at layers 4-5. But no amount of investment at the bottom produces the cognitive layer where transformation actually happens. You can't make task workers incrementally smarter until one of them figures out how to be the VP.

The cognitive layer (layer 2) is where the invisible 80% of knowledge work operates. It's the governed cognitive extension that holds context, translates intent into action, and makes everything below invisible. That's the layer nobody has figured out yet.

Maps to organizational structure

The cognitive stack mirrors how organizations already work. Corporate org charts have more thinking at the top and more doing at the bottom. Work starts with intent and flows down. This isn't a new idea—it's the shape intelligence naturally takes when coordinating complex work.

August 28, 2026 update—three independent systems already prove this, not just an organizational analogy. Cursor's agents, StrongDM's Software Factory, and Anthropic's own agent teams all independently converged on hierarchical management structures without being designed to. Hierarchy isn't a human organizational preference—it's what intelligence does when it has to coordinate complex work at scale, and now there's real evidence beyond the org-chart parallel.

Using this framework

Deploy when:

• Evaluating AI investments: ask "which layer are you buying?" Most products live at layers 4-5
• The "agents" conversation needs reframing: agents are claws (layers 4-5), not the brain (layer 2). The transformation is at the top, not the bottom
• Someone asks why second brains work: they start at layer 2 where context and judgment live, then reach down into systems as needed. Agents become invisible infrastructure
• Challenging bottom-up AI strategies: "you'll never transform your way to the future by deploying increasingly smarter agents"
• Explaining why corporate AI initiatives stall: they're optimizing layers 4-5 while the cognitive layer (layer 2) remains unbuilt
• The pitch is about agent orchestration: redirect to the question of who builds the brain layer—the governed cognitive extension where context lives and intent gets translated into action