Overview
Introspection runs the agents you build, records what they do, and helps you turn that record into a better agent.
Introspection is for teams that ship an AI agent to their own customers and need to improve it from real use. You define the agent in Git as a recipe; Introspection deploys it, runs every task in its own sandbox, and keeps the full record.
You build a recipe and prove it on your machine, ship it as a versioned runtime, run tasks from your app, Slack, or a schedule, observe what happened, and improve it with evals and experiments. A coding agent such as Claude Code can drive this through the CLI (Using a coding agent).
Where to go
| You want to | Start here |
|---|---|
| Deploy a first agent | Quickstart |
| Learn the terms | Concepts |
| Solve a specific problem (Slack, schedules, files, your app) | How-to |
| Write or port an agent | Build |
| Release and roll back | Ship |
| Find out why an answer was wrong | Debug a bad answer |
| Make the agent measurably better | Improve |
| Answer a security review | Security |
Last updated on