Building agents that recover from failure

Agents fail constantly — bad tool calls, hallucinated arguments, flaky APIs. The difference between a demo and a product is what happens next.

Validation gates

Before an agent's output reaches the user (or the next step), validate it against a schema or a set of hard constraints. Cheap checks catch most failures before they compound.

Retry with context, not blind repetition

A naive retry just re-runs the same prompt and gets the same mistake. A useful retry feeds the failure back in: what was tried, what broke, and what constraint was violated.

attempt 1: tool_call(args) -> error: missing required field "id"
attempt 2 prompt: "Previous call failed: missing 'id'. Retry with a valid id."

Knowing when to stop

An agent that retries forever is worse than one that fails fast. Cap retries, and when the cap is hit, surface a clear, specific error instead of a vague failure — that's what actually saves debugging time later.