a16z on Agent Loops: The Art of Knowing When to Stop

stuffyokodraws · x · 2026-08-07

Yoko Li from a16z explores a core challenge in AI agent loop engineering: how does a model know when its work is done?

The article points out that an AI model can theoretically revise paragraphs, try new code implementations, or generate images indefinitely. It doesn't get tired, nor does it notice that the last three revisions made the result different but not necessarily better. Humans stop working not because they reach an objective state of perfection, but by relying on external signals—like a programmer waiting for tests to pass or PR approval, a designer accepting minor flaws after weighing trade-offs, or a writer hitting a deadline.

Thus, "done" is rarely a property of the work itself. It is a judgment produced by the system around the work, utilizing signals like tests, specs, approvals, and deadlines. To build effective autonomous agent loops, developers must construct similar external validation signals and stopping mechanisms for the AI.

Related event: a16z Explores Termination Conditions for AI Agents(3 posts)→

Original post →

More from coding & agent

coding & agent channel →