An invoice was delivered, but the acknowledgment never arrived. A retry can create a second delivery even though the final response sounds perfectly reasonable.
Agent Crash Lab is a ufodia flight simulator for action-taking workflows. It places durable state beside an event timeline, so you can inspect what committed and compare a deliberately naive policy with a guarded one.
Rehearse the missing receipt
Open the lab, select The missing receipt, and choose Run experiment. Watch how the two policies handle a timeout after a commit.
Pause or step through the events. Compare the delivery ledger and operation IDs with the messages each policy received.
Try a duplicate task, partial batch, stale read, or late response. Export the traces to keep an inspectable record of the modeled behavior.
The included experiments run locally without an account, API key, or model call. Nothing is delivered to a real account.
What the experiment establishes
These are small, deterministic simulations with explicit service assumptions. They do not run live LLM agents, measure model quality, or certify a production integration. Imported traces are structurally validated observations, not independently verified evidence. Real APIs still need testing against their own contracts.
Open Agent Crash Lab · Inspect the MIT-licensed source
If this helps your team reason through recovery, one optional coffee supports ufodia's examples, tests, and documentation.
Real Agent Crash Lab workspace showing the synthetic missing-receipt experiment. Both policies are deterministic simulations; no invoice is sent.
