Your agent's retry bug isn't an AI problem. It's a distributed-systems problem with an LLM on top.

Most agent demos have a happy path: the model picks a tool, the tool returns 200, the agent moves on....