What is a rollback?

A rollback is the act of returning a system to its previous, known good version after a change causes problems. For an AI agent, rolling back means restoring the earlier brief, sources, model, tools and bounds together, and dealing separately with any actions the faulty version already took.

Why does a rollback matter for agents?

Agents change often, and not every change behaves on live work the way it did in testing. A fast, rehearsed rollback limits how long a bad version can act. Without one, teams try to fix forward under pressure, making further changes to a system they do not yet understand.

The trap is assuming a rollback undoes the damage. It does not. It stops the faulty version from doing more. Letters already sent, records already changed and payments already prepared remain, and each needs its own decision.

What does rolling back an agent involve?

An agent is more than code. A full rollback restores:

• the brief and instructions;
• the source configuration, including which documents and systems it reads;
• the model and version;
• tool connections and the identity's permissions;
• customer limits, if they changed with the release.

Restore them together. Rolling back the brief while leaving the new source in place produces a third version nobody tested. The previous version's certificate also needs checking: if it has lapsed in the meantime, the restored agent should act no further than its current evidence supports.

An AI agent rollback plan template lists what to record before each change so the rollback is ready.

What does an example look like?

A US insurer releases a new version of its claims intake agent, adding a source for state-specific notice requirements. Within a day, handlers see the agent asking customers in some states for documents that are not needed. The operations lead triggers the rollback: the previous brief and source set are restored together, and the agent is back on the earlier version within minutes. Then the team turns to the actions already taken: the verdict record lists every customer asked for an unnecessary document, and a handler writes to each to withdraw the request. The setting is described in AI agents in US insurance operations.

How is a rollback different from a reversal?

A rollback changes the agent. A reversal changes the outcome of an action, such as cancelling a payment or correcting a record. They are separate decisions with separate owners. Rollbacks belong to whoever runs the agent. Reversals belong to the business owner of the affected process and follow that process's normal controls; some, such as reversing a settled payment, should never be attempted by an agent at all.

It also differs from a kill switch. A kill switch stops the agent acting, without changing its version. A rollback keeps the agent acting, on an older version. In an incident, teams often pull the switch first, then roll back once they know the earlier version is safe.

When should you roll back rather than fix forward?

Roll back when the earlier version is known to be good and the cause of the problem is not yet clear. Fix forward only when the fix is small, understood and testable quickly. In wealth management, for instance, where an agent drafts client communications subject to suitability rules, rolling back to the last approved version is nearly always safer than an untested patch; the setting is set out in AI agents in US wealth management operations.

Which terms sit close to it?

Blue-green deployment makes rollback a single switch. Canary release limits what needs rolling back. Change management records what changed, which is what the rollback restores. See our guide to deploying AI agents for the wider release process.