An agent after failure
Recovery, retries and idempotency through a CI example
Slide contents
1. An agent after failure
Recovery, retries and idempotency through a CI example
2. Recovery needs several distinct guarantees
Resume execution, establish the outcome, repeat safely
3. Experience does not confirm operation completion
Continuing the memory deck and episode 42 on evals
4. A timeout leaves the outcome unknown
Atlas · commit c3 · Node 22 · verification request op-17
5. An empty read does not prove absence
Three hidden outcomes behind the same timeout
6. LIMBO compares actions in matched worlds
Jiapeng Li · Microsoft · arXiv v1 · September 24, 2026
7. An error code hides the side effect
Lost responses, late commits and duplicate delivery
8. Compensation does not erase a duplicate
Final state and effect history are checked separately
9. Late effects change the failure pattern
E1 · six frontier models · episodes with duplicates
10. Keys help when the service honors them
E2 / E2k · four models · duplicates, %
11. Waiting depends on delay bounds
E5 · delays of 40 s–2 h · simulated time
12. Contract effects have an experimental scope
Simulated services, a shared gateway and restricted tools
13. Recovery needs both conversation and files
Moe Abadi · Temporal · October 8, 2026
14. An intent is not a confirmed result
A claim before invocation; a result after return
15. An old worker may finish later
Revoking journal writes does not fence an external API
16. A key identifies intent, not arguments
Retrying op-17 and starting op-18 can share one SHA
17. The key and effect must commit together
A local marker cannot make an external API transactional
18. Retries across layers multiply load
Five layers with three attempts each: up to 243 calls
19. Jitter spreads synchronized retries over time
Bound attempts, increase waits, spread clients
20. Unknown outcomes need their own branch
Recovering op-17 after losing the response
21. Test effects at failure boundaries
Matched tasks and faults for compared policies
22. Resuming work does not guarantee one effect
Executor state, service contract and outcome evidence
23. Uncertainty must survive a restart
What was intended, what happened, what can be repeated?