Exponential Backoff - Tools Against Typed APIs, with MCP as the Wrapper.

Follow me on Linkedin:   / aziz-ketata-08ba20158   After this episode the viewer can look at any error a tool returns and say retry safe, stop, or unknown, and can write down the wait schedule, the budget and the last message. Retries are where agents do damage because the retry decision is made in three places, the provider SDK, the agent loop and the HTTP client, and none of them sees the others. The classification depends on two things: whether the request may have reached the server, and whether the operation is idempotent. A read timeout after the bytes were sent is the unknown case, and an unknown write is never re-issued blindly. The wait schedule is jittered exponential backoff with a finite budget, and the last message the model sees is an instruction, so it says what happened, what was not done, and what the model may do next.