Part 1 of 6 · 1 min read
Overview
In this module you will
- Say what an agent's state is and keep it as a fold over an event log.
- Record a trace at a chosen level of detail, and know what each level lets you answer.
- Replay a run exactly from its trace, and test a new prompt on old tool results without calling any tool.
- Resume a run after a crash without sending the same email twice.
- Keep secrets and personal data out of traces.
- Read a trace as a timeline that a stranger could follow.
When an ordinary program fails, you read a stack trace and re-run it. An agent is different on both counts: the cause of a failure is usually a decision (the model said something, a tool returned something) that is gone once the run ends, and re-running it does not reproduce it, because the model is random and the world has moved on. The only record of what happened is what you stored while it ran. Deciding what to store is therefore a design decision made before the failure.
About the examples
Everything here is an offline imitation. The "model" is a scripted program that decides from the last message and makes one noisy decision (it sometimes quotes a stale date). The "world" is a tiny order desk with a carrier whose answer drifts each time it is asked, and an outbox. The numbers show what each kind of record buys you, not what your production traces look like. No real model or service was called, and no real email was sent.