ALT Two-column comparison. Left column titled Sign with three boxes labelled Agent has production credentials, A single API call destroys production, and System prompt says don't, ending in a red badge labelled production deleted. Right column titled Control with three boxes labelled Production credentials in CI/CD secrets only, Production changes flow through the pipeline, and Pipeline gate scores commit push and release, ending in a green badge labelled denied at the boundary.

ALT Screenshot of a Claude Code terminal session showing an agentic risk-gating workflow. A risk-scorer tool rates a pending push as 6/25 (Medium), which exceeds the configured appetite threshold of 5. When Claude attempts to push via npm run push:watch, a PreToolUse Bash hook blocks the action with the error "Push risk score 6/25 (Medium or above). Reduce unpushed changes first." Claude then explains the score and plans to add an integration test covering the full event-to-cache-to-API pipeline to reduce the risk score before retrying.

ALT Process flow diagram showing how the architect gate works: prompt starts the turn, AI attempts an edit, a gate check validates the marker, and if no marker exists the edit is denied until an architect agent reviews and passes.