# What is Introspection?

Introspection is the infrastructure for long-horizon vertical agents: build them, run them, and improve them while owning the behavior yourself.

Agent behavior lives in a **recipe** — the open format for vertical agents, built on Pi — versioned in Git. In an Introspection **runtime**, Pi runs the agent loop from a pinned recipe version inside an isolated sandbox. Introspection records what happened and gives you the evidence to improve it deliberately.

```text
Recipe → Runtime → Production evidence → Better recipe
  ↑                                           │
  └───────────────────────────────────────────┘
```

Read more about [Pi, recipes, and how they relate to Introspection](/recipes).

- **Recipes** keep instructions, skills, tools, policies, and quality criteria together in Git.
- **Runtimes** run Pi with pinned recipes, scoped identity, credentials, and infrastructure.
- **The learning loop** turns completed work into observations, patterns, evaluation, and focused recipe changes.

The model is interchangeable. Your definition of good is not — it lives in the recipe and its quality criteria, and it improves specifically for your users.

## A concrete example

A support agent begins mishandling refund exceptions. Instead of adding a vague warning to its prompt, you inspect the source conversations and discover a recurring missing escalation procedure. You update the associated recipe skill, check the behavior locally, and add a judge if that risk deserves ongoing measurement. The resulting change remains reviewable and reproducible in Git.

## What lives where

| Layer | Responsibility |
| --- | --- |
| **Pi** | The agent runtime that runs the agent loop. |
| **Pi Recipes** | The open format for vertical agents, built on Pi, and its toolchain. |
| **Introspection Platform** | Runtimes, task evidence, evaluation, experiments, security, and operations. |

## Continue

There are two equally supported ways in. If you work with a coding agent, the
[**Quickstart**](/start/quickstart) drives creation and migration. If you prefer the terminal, start from the
[**agent development lifecycle**](/guides/development-lifecycle) and the
[local toolchain](/recipes/cli). Both produce the same recipe.

- Use the [**Quickstart**](/start/quickstart) to build, run, and deploy your first agent.
- Follow the [**agent development lifecycle**](/guides/development-lifecycle) to author a recipe and move each later change to production.
- [**Work with repositories**](/guides/work-with-repositories) when the agent needs a checkout of your source code.
- [**Learn from production**](/guides/learn-from-production) once real traffic starts telling you what to fix.
- [**See how Introspection works**](/platform#how-introspection-works) by following one piece of work through the platform.

What is Introspection?

Introspection is the infrastructure for long-horizon vertical agents: build them, run them, and improve them while owning the behavior yourself.

Agent behavior lives in a recipe — the open format for vertical agents, built on Pi — versioned in Git. In an Introspection runtime, Pi runs the agent loop from a pinned recipe version inside an isolated sandbox. Introspection records what happened and gives you the evidence to improve it deliberately.

Recipe flows to Runtime, then Production evidence, then Better recipe, which loops back to Recipe — Text representation of the diagram:
`Recipe → Runtime → Production evidence → Better recipe ↑ │ └───────────────────────────────────────────┘`

Recipes keep instructions, skills, tools, policies, and quality criteria together in Git.
Runtimes run Pi with pinned recipes, scoped identity, credentials, and infrastructure.
The learning loop turns completed work into observations, patterns, evaluation, and focused recipe changes.

The model is interchangeable. Your definition of good is not — it lives in the recipe and its quality criteria, and it improves specifically for your users.

A concrete example

A support agent begins mishandling refund exceptions. Instead of adding a vague warning to its prompt, you inspect the source conversations and discover a recurring missing escalation procedure. You update the associated recipe skill, check the behavior locally, and add a judge if that risk deserves ongoing measurement. The resulting change remains reviewable and reproducible in Git.

What lives where

Layer	Responsibility
Pi	The agent runtime that runs the agent loop.
Pi Recipes	The open format for vertical agents, built on Pi, and its toolchain.
Introspection Platform	Runtimes, task evidence, evaluation, experiments, security, and operations.

Continue

There are two equally supported ways in. If you work with a coding agent, the Quickstart drives creation and migration. If you prefer the terminal, start from the agent development lifecycle and the local toolchain. Both produce the same recipe.

Use the Quickstart to build, run, and deploy your first agent.
Follow the agent development lifecycle to author a recipe and move each later change to production.
Work with repositories when the agent needs a checkout of your source code.
Learn from production once real traffic starts telling you what to fix.
See how Introspection works by following one piece of work through the platform.