Delegation, not automation

Workers don't think like programmers. They think like managers. They don't want to build workflows—they want to hand off tasks. The entire enterprise AI industry is investing at the wrong layer of the stack.

Published: December 18, 2025 — Original post Extended: February 25, 2026 — The Cognitive Stack formalizes the skills hierarchy as a named five-layer framework, adds Karpathy's "claws" framing, and maps two industry trajectories (bottom-up automation, top-down AI) colliding in the middle.

The automation fallacy

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.

The result is always the same: 1% of workers who think like programmers build amazing things. The other 99% keep doing their jobs the way they always have.

This isn't a failure of the tools. It's a failure of the premise. Knowledge workers don't have stable, repeatable workflows worth the investment 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 delegation model

What workers actually want is what every manager has always wanted: a capable entity they can talk to, that understands context, and that handles things. Not "build an automation for expense approvals." Just: "approve Brian's expenses."

The difference:

The skills hierarchy

The delegation model reveals a hierarchy that the automation-obsessed industry is investing in upside down:

Worker (states intent, exercises judgment)
  └─ Cognitive Extension Agent (the "brain"—what the worker talks to)
       └─ Skills (what the brain knows how to do)
            └─ Agentic sub-processes (how skills get executed)
                 └─ Interfaces (one possible execution tool)

The value is at the top. The industry keeps investing at the bottom.

• Worker: "I need this report packaged and sent to the whole team before Friday."
• Cognitive extension: Knows who's on the team, knows the context, knows how you like materials packaged.
• Skills: Knows how to compile documents, format for distribution, check for sensitive content.
• Agentic sub-processes: Opens a browser to access shared drives. Calls an API to check calendars. Drafts emails.
• Interfaces: Maybe it triggers an approval step via API. Maybe it uses a Slack connector. Maybe it calls an MCP server. These are implementation details the worker never sees or cares about.

The worker talks to the top of the stack. Everything below is invisible. That's the point.

Where agents actually fit

"Agents" as the enterprise AI industry uses the term are mostly workflow automation bots—the bottom of the hierarchy. They automate specific tasks within specific systems. They're RPA with better NLP.

The cognitive extension agent at the top of the hierarchy is fundamentally different. It's not automating a workflow. It's extending a human's cognitive capacity. It has persistent context, evolving knowledge, judgment amplification. It's the thing you talk to.

Both types of agents matter. But the industry conversation is stuck on the bottom-of-stack agents (workflow bots) when the transformation is happening at the top (cognitive extension). The workflow bots become tools the cognitive agent uses, not things workers interact with directly.

Agents aren't the solution. Agents are a form factor. They're an implementation detail of the delegation model—one of several ways a cognitive extension gets things done. Nobody cares whether the AI used an API call, a browser agent, or a computer-using agent to check their expense report. They care that it got checked.

The iPhone analogy

Today's AI automation studios are the equivalent of building BlackBerry apps. Powerful, useful for the 1% who invest the time, and completely 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 so intuitive your grandparents could use it. AI will get there too. Workers won't build automations. They'll just talk to their AI, and if it needs to fire up an agent, open a browser, call an API, or trigger a workflow, it will. In the background. Invisibly.

The automation studios being built today aren't wasted effort. They're the plumbing—the connectors, the access patterns, the API integrations—that future cognitive agents will use. But the studios themselves are stepping stones, not the destination.

Using this framework

Deploy when:

• An enterprise is investing heavily in "automation studios" expecting rank-and-file adoption—workers won't build, they'll delegate
• The "agents" conversation is stuck on workflow bots—redirect to the hierarchy: agents serve the cognitive layer, not the other way around
• Someone asks "did second brains skip agents?"—no, agents are still there, they're just in the background where they belong
• Product strategy is focused on the bottom of the stack (connectors, workflow designers, orchestration engines) instead of the top (what the worker actually talks to)
• The pitch is "workers will automate their own jobs"—the 20-year track record of RPA, low-code, and citizen development says otherwise

The hierarchy diagram is the key artifact. Draw it. Point to where the industry is investing (bottom). Point to where value accrues (top). The gap between those two points is where the strategic opportunity lives.