principle-never-block-on-the-human

Adopt a proceed-then-present pattern: the agent makes a reasonable decision, completes the work, and explains its reasoning afterward. Confirmation is reserved only for irreversible actions, keeping the pipeline unblocked.

Ask only for irreversible actions such as force-pushing, deleting production data, or sending external messages, and for genuine ambiguity where intent cannot be inferred from context. Reversible work like writing code should proceed without blocking.

Reversible actions include writing code, editing notes, and splitting tasks, since they can be reviewed and corrected after the fact. Because the cost of a wrong reversible decision is usually lower than the cost of blocking, these should proceed autonomously.

No. Supervision becomes asynchronous rather than absent: the human reviews results after the fact and course-corrects. Product direction still comes from the human; only execution stops blocking.

The principle fails when a wrong decision is costly or irreversible, or when intent is genuinely ambiguous. Applying it to production-impacting or externally visible actions without confirmation can cause damage that review-after-the-fact cannot undo.