Retrying sounds simple until the first operation might already have committed. A longer timeout can help completion while making latency worse. A repeated write can look successful while duplicating an effect. Agent Rehearsal lets you compare recovery policies against paired, deterministic fault samples. Inspect each attempt, duplicate writes, call budgets and declared virtual latency. The Python example reproduces a lost payment response with in-memory functions and checks that a guarded retry charges once. There is also a local trace inspector and a zero-dependency Python SDK. The simulator is a fixed workflow, not an LLM benchmark, and its percentages should not be read as a model leaderboard. The source is MIT licensed; the public simulator needs no account, API key or payment. I am looking for feedback on which failure details need to be visible before a recovery policy feels understandable. Try it: https://agent-rehearsal.web.app Source: https://github.com/shi1720/agent... Built by Shivam Gupta. More work and contact: https://shivamgupta.web.app/ LinkedIn: https://www.linkedin.com/in/shiv...

Agent Rehearsal

Rehearse tool failures before retries reach production

Rehearse tool failures before retries reach production