Four production AI agents crashed mid-action: none recorded what already finished

merlinofthewater · reddit · 2026-10-09

The author ran kill-and-retry experiments on four published open-source agent codebases, killing each mid-action and retrying like a supervisor or worker pool would, with a ledger from a fake service that fsyncs before responding.

Cua (computer-use): killed mid-checkout, the retry charges the card again — every run, on both 0.8.4 and 0.9.0.
Skyvern: any process death strands the task in running forever — not failed, no alert fires; the only exit is a rerun, which orders again.
Corsair: two replicas consume the same rotating OAuth refresh token; a crash permanently breaks stored credentials.
Inconvo: an interrupted answer costs 60% more model calls because the checkpointer logs the graph node but not the fifteen calls beneath it.

Shared root cause: none records which actions already completed, so retries start with empty history and redo work.

Citing arXiv 2609.29095: across 25,930 agent episodes graded against a ledger of committed effects, agents self-reported success in 90% of episodes where they had duplicated an effect — benchmarks relying on self-reported traces score "duplicated and lied" as a pass.
Unfixable boundary admitted: death between action landing and record committing still repeats; effect and record live in different systems.
Practical takeaway: agents touching money or side effects need external ledgers/idempotency keys, not agent self-reports.
Disclosure: the author builds a runtime for this problem; harnesses and raw ledgers are public and need no API key.

Original post →

Four production AI agents crashed mid-action: none recorded what already finished

More from coding & agent