• Sources: primary
  • Summary: Tanmay Sah and Dolly Sah report 2,880 paired trials in which agents completed 84% of tasks normally, but only 47% of normally successful trials recovered after an injected failure, and naive retries after a lost acknowledgement produced duplicate effects in 54% of such trials. In their EvoUndo work, 197 of 478 harness changes that improved capability left the system unable to recover, prompted repair fixed none of them, and capturing a pre-state witness plus a richer undo language recovered 180 of 197. Requiring both capability gain and verified recoverability raised admissible improvements from 281 to 461 of 600 tasks.
  • Why it matters: Agents that touch external APIs, configuration or their own harness need state capture, idempotency checks and verified compensation, since git revert covers only the repository.
  • Background: A timeout tells a client that the response never arrived, not that the operation failed, which is why distributed systems use idempotency keys and compensating actions instead of blind retries. The same problem now applies to coding agents that create tickets, tags or deployments. Self-evolving agents add a second layer, because they also rewrite their own prompts, tools and middleware.