Give Coding Agents a Tested Way to Undo Their Work

Tanmay Sah and Dolly Sah present their UndoBench and EvoUndo research on recovery after coding-agent failures.

A timeout is not necessarily a failed action

An external service may create a ticket even if its acknowledgement never reaches the agent. Retrying blindly can create another ticket.
Git can restore repository changes, but runtime configuration, database writes, deployments and notifications can require separate recovery steps.
The response depends on what actually happened: retry, restore saved state, compensate for an external effect, or stop and escalate.

EvoUndo captures evidence of the original state before a change and tests a recovery plan in a test environment.
An independent verifier checks the resulting state; running an undo command alone is not proof that recovery succeeded.
The speakers report that 197 of 478 capability-improving changes in their experiment had recovery defects, about 42%. This is a result for their tested setup, not a universal failure rate.
They describe improved recovery after providing richer recovery operations. These are research results, not proof that every real-world side effect can be reversed.

Treat recoverability as an acceptance condition for agent changes, alongside the capability they add.
Save the original values and test the actual recovery procedure before relying on it in production.
Irreversible or unverifiable actions still need a human escalation path.