H2 AI Labs
Lab Notes

Sep 18, 2026 · 6 min

Why AI Agents Need Verification, Not Just Actions

An action is only useful when the intended outcome can be confirmed.

This note is part of an evolving builder notebook. The central question is not whether AI can produce an action, but whether the larger system can make that action understandable, controllable, and verifiable.

The working principle

AI output should expose its assumptions and evidence. Deterministic software should own rules that must remain stable. Humans should retain meaningful control whenever consequences matter.

What changes in practice

This shifts design away from a single impressive model response and toward explicit state, observable workflows, correction paths, and outcome checks. The result is less magical—and far more useful.

Related experiment: Loose Ends