Rolling back an agent takes three moves, in this order: drop its depth straight away so it can only read, restore the last certified version of its brief, sources, model and bounds, then review every action it took during the bad period. Reverting the configuration is the easy part. Finding and correcting what it already did is the real work.
What exactly are you rolling back?
An agent is not a single file. Its behaviour comes from a bundle: the brief that describes its role, the sources it may read, the connections it may use, the model behind it, its bounds and its certificate. A bad change can sit in any of these. A rollback restores the whole bundle to the last version that was certified, not just the piece you suspect.
That only works if every bundle is versioned and the certified one is marked. If your team edits the brief in place, there is nothing to roll back to. The playbook on rolling back an AI agent after a bad change covers how to set up versioning before you need it.
What are the steps, and who takes each one?
- On-call owner: drops the agent to depth 1. It can still read and report, but cannot draft for sending or act. This takes seconds and stops further harm. Evidence: a filed verdict naming who lowered the depth and why.
- Operations lead: routes incoming work to people. Queues that the agent was clearing go to named staff. Evidence: the time the routing changed and the queue handed over.
- Platform owner: restores the last certified bundle. Evidence: the version restored and the version it replaced.
- Analyst: pulls every action from the bad period from the filed record, using the time the change went live as the start. Evidence: the list of affected cases.
- Process owner: decides what to correct. Some actions are fine, some need reversing, some need a customer contact. Evidence: a decision per case.
- Owner: recertifies before restoring depth. The restored bundle was certified once, but the certificate may have lapsed, and the incident may have revealed a gap. Evidence: a fresh certification result.
What can the agent do during a rollback, and what stays with a person?
| Task | The agent | A person |
|---|---|---|
| Listing affected cases from the record | Yes, at depth 1 | Checks the list is complete |
| Drafting correction letters | Yes, at depth 2, once restored and checked | Approves and sends |
| Reversing actions | No | Decides and acts, case by case |
| Deciding which actions were wrong | No | Process owner |
| Restoring depth | No | Owner, on new certification |
Some corrections may not be possible at all. Reversing a payment that has settled on a rail, for example, is an action the platform may withhold for every agent, so a person must arrange recovery through the proper route.
What does a rollback look like in practice?
Say a lender changes the brief for its arrears agent to soften the tone of reminder letters. The edit also removes a line telling the agent to check for an active payment plan. Over two days at depth 3, it sends 120 reminders, 15 of them to customers who were already on a plan.
A complaint arrives on the third morning. The on-call owner drops the agent to depth 1 within minutes and the collections team takes over the morning's queue. The platform owner restores the previous brief. The analyst filters the record for every reminder since the change and matches it against active plans, which finds the 15. The owner approves an apology letter to each, drafted by the restored agent and sent by a person, and the arrears notes are corrected. The agent goes back through certification with a new test case for customers on plans.
Which controls make rollback possible?
- Versioned bundles with the certified version marked.
- A record of every action, so the bad period can be reconstructed exactly.
- A tested route for people to take work back. Our view is that handing work back from an agent must always be possible, and the steps are in taking work back from an AI agent.
- Small, staged changes. A change that went to a slice of work first is much easier to undo, which is the case for rolling out an agent gradually.
The wider sequence is in our guide to deploying an AI agent.
Where Pryme fits
In Pryme, every verdict is filed in an append-only record that names the bound which decided it, so the bad period can be reconstructed action by action. Lowering a capability to depth 1 is immediate and itself filed. Restoring depth needs Attest to certify the restored version on your cases; the earlier certificate does not carry it back.
Questions people ask
Who should be allowed to trigger a rollback?
The process owner and whoever is on call for the agent, without needing anyone else's approval. Restoring depth afterwards should need more: fresh certification and the owner's sign-off.
How fast should a rollback be?
Lowering depth should take seconds. Restoring the bundle should take minutes. Reviewing past actions takes as long as it takes, and it should not be rushed to get the agent back.
Should we tell customers about a rollback?
Tell the customers whose outcomes were affected, with a correction. A rollback itself is an internal control, but its consequences may need explaining.