The Work Primitive: What Every AI Product Leader Gets Wrong

A semantic work primitive is a unit of work — such as a refund, a reschedule, a payment authorization, or a compliance exception — that carries its full meaning, permissions, implications, and reversibility in a way an agent can read and act on (5:30). This is contrasted with agents simply filling in database fields or clicking buttons without understanding the human intent behind those actions. The concept is central to the speaker's thesis that durable AI products must expose these primitives, not just UI access.

The speaker introduces a hierarchy of three layers agents must navigate (2:45). Access is the ability to reach tools and interfaces (computer use, browsers, MCP). Meaning is understanding what an object or action represents, who owns it, and what its consequences are. Authority is who is allowed to do what, under what conditions. Most current AI products only address the access layer, leaving meaning and authority as unsolved product challenges.

The speaker argues (1:30) that while computer use (e.g., OpenAI Codex, Claude) is necessary as a universal adapter for legacy software built for humans, it is inherently a shallow interface. A screenshot shows the agent what's on screen but doesn't reveal domain structure; a browser reaches web apps but doesn't know whether a button is reversible or financially material. Computer use is the bridge to the old world, not the long-term competitive advantage.

Coding agents succeeded first not merely because code is text and LLMs are good at text, but because software development already possesses unusually rich work semantics (10:00) — modules, dependencies, tests, type systems, linters, git history. This semantic density gives agents a feedback loop (run test, see error, revise) without requiring constant human supervision, making code a natural wedge for agentic work.