Methodology

Escape Hatch

Predefined conditions in a PRD that trigger automated execution to stop and request human input instead of continuing blind.

Escape hatches prevent agents from going off the rails during autonomous execution. Each PRD phase defines conditions that should halt progress: "if the API returns an error on 3 consecutive attempts, STOP", "if the output file exceeds 500 lines, STOP", "if the confidence score drops below 6/10, STOP". Without escape hatches, Ralph Execute would continue iterating on a broken premise, burning tokens and producing garbage. The BLOCKED_NEED_HUMAN signal in Ralph is the escape hatch in action. Escape hatches were formalized after an agent spent 14 iterations optimizing a prompt that had a typo in the system instruction.

Where it shows up:

PRD designAgent autonomy boundariesRalph executionCost control

Build your Knowledge OS

90 minutes from zero to your first skill chain.