An agent acts in a second. Undoing what it got wrong can take days — or be impossible.
Keel compiles a verified way to reverse every action before production, holds back what can't be undone, and proves what happened — inside your systems of record.
Permissions decide what an agent is allowed to do. Nothing decides whether you can take it back.
A model fires a write in milliseconds. The action that reverses it may not exist at all.
Knowing what happened doesn't reverse it. Audit trails record the damage; they don't restore.
An agent with access can do real, final harm entirely within its granted rights.
Evals predict behavior in a sandbox. They don't stop the action once it's live.
— so the keel brings it back, before it runs.
A tested undo is compiled and stored beside the action. It runs freely, with the rollback held in reserve.
High-impact and paused. Routed to a named human with the full blast radius attached before it can execute.
No compensating action exists. Keel blocks it from running blind — the agent can't quietly route around it.
Simulate the agent against real exceptions and compile a tested rollback for every write.
Stop what can't be reversed. Irreversible actions route to a human, never autopilot.
Hold the same line live. Reversible flows, risky holds, drift pauses the agent.
Hash-chained evidence of every verdict, undo, and approval — audit-ready.
Wrap the agents you've already built — Keel reads intended actions before they execute.
Compile and enforce rollbacks where finality actually lives — ledgers, ERPs, payment rails.