Authorise the irreversible
RAPID's dividing line between "agents build" and "engineers govern" is not a feeling — it's an enumerated list. Agents investigate, build, verify and capture. A named human authorises anything that can't be cleanly undone, regardless of how confident the agent is.
This is the crisp boundary that makes it safe to let agents move fast: everything reversible is theirs; everything irreversible is ours.
The rule
An action on this list requires explicit authorisation from a named human, recorded, before it runs. Agent confidence is never a substitute. "Recorded" means the authorisation is captured where it can be audited — in the PR, the release evidence pack, or a register — with who authorised, when, and on what evidence.
The actions
| Action | Who authorises | Evidence they check | Recorded in |
|---|---|---|---|
| Production deploy / release | DevOps or Lead (per release) | Production Readiness gate green | Release evidence pack |
| Schema or data migration | Senior Dev + Data | Reversible/rollback plan; backup verified | PR + migration record |
| Data deletion (incl. bulk/irreversible purges) | Data owner | Scope, legal basis, retention rules | Change record |
| Key / secret rotation | Security | Blast radius; rollout plan | Security register |
| Client-facing communications | Commercial / account owner | Approved narrative | Comms record |
| Major dependency upgrade | Senior Dev | Changelog analysis; test evidence | PR + ADR if flagged |
| Spend commitment (tooling, tokens beyond budget, infra) | Budget owner | Cost, funding source | Finance/register |
| Auth / access-control changes | Security | Least-privilege review | Security register |
| Standing configuration (mail rules, webhooks, integrations, recovery contacts) | Service owner | Purpose; reversibility | Change record |
| Public / external publishing | Owner per content type | Content review | Publish record |
This list is a floor, not a ceiling. A squad may designate additional irreversible actions for its product; it may not remove one.
How agents treat it
- Agents are configured to stop and request authorisation at these points rather than proceed — enforced by the agent guardrail standard and the gate cards, not left to prompt discretion.
- An agent may fully prepare an irreversible action (write the migration, draft the comms, stage the deploy) — preparation is reversible. Only the irreversible step waits for a human.
- The authorisation is per-action. Approving one deploy does not pre-approve the next.
Why this is the whole game
Reversible mistakes are cheap when the loop is cheap — you fix them and deposit a gotcha. Irreversible mistakes are not. Concentrating human judgement exactly on the irreversible is what lets everything else run at agent speed.