# What is Introspection?

Introspection is the platform for building vertical agents, running them for every customer, and distilling production experience back into their agent recipes.

Agent behavior lives in an **agent recipe**: versioned source in Git that Pi can
run locally or inside an Introspection runtime. The platform supplies governed
execution, customer isolation, durable evidence, evaluation, and controlled
deployment without taking ownership of the behavior you built.

## Start with the job you have

Build an agent recipe
Define one useful agent, test it on your computer with Pi, and create its first runtime.
Deploy it for customers
Connect your local recipe in development. Then deploy the reviewed version to staging and verify the production version.
Add it to your product
Start tasks from your application, stream their output, and tie each task to a user or customer. You can also reconnect to a task and save its results.
Learn from production
Turn real usage into signals, validate the smallest useful change, and distill it back into the recipe.

## How the system fits together

```text
Agent recipe → Runtime → Production experience → Signals
     ↑                                            │
     └──── New recipe version ← Validated learning ┘
```

| Layer | Responsibility |
| --- | --- |
| **Agent recipe** | The agent, its instructions, skills, tools, and quality criteria in source you own. |
| **Pi** | Loads the recipe and runs the agent on your computer or in a managed runtime. |
| **Introspection runtime** | Runs one fixed recipe version. It handles customer work in separate sandboxes and saves the results. |
| **Continual learning loop** | Collects signals from real use, tests proposed changes, and adds approved changes to the next recipe version. |

The model is interchangeable. Your definition of good is not: each validated
lesson returns to the agent recipe, where production experience compounds into
durable domain IP.

## A concrete example

A refund agent promises an exception for a final-sale order. Feedback and
observations show the problem, and related conversations show that it has
happened more than once. You identify the missing policy decision and add
representative test cases. You then test a focused change that fixes the problem
without breaking ordinary refund requests. The approved change goes into the
recipe as a reviewable Git change.

## Choose your next page

- If you are new to Introspection, complete the [Quickstart](/start/quickstart).
- If you need definitions for product terms, read [Core concepts](/start/concepts).
- If you already run an agent, begin with [Learn from production](/guides/learn-from-production).
- If you are building an application, choose an [SDK](/sdk).
- If you need a command, use the [CLI reference](/sdk/cli).

What is Introspection?

Introspection is the platform for building vertical agents, running them for every customer, and distilling production experience back into their agent recipes.

Agent behavior lives in an agent recipe: versioned source in Git that Pi can run locally or inside an Introspection runtime. The platform supplies governed execution, customer isolation, durable evidence, evaluation, and controlled deployment without taking ownership of the behavior you built.

Start with the job you have

Build an agent recipeDefine one useful agent, test it on your computer with Pi, and create its first runtime.Deploy it for customersConnect your local recipe in development. Then deploy the reviewed version to staging and verify the production version.Add it to your productStart tasks from your application, stream their output, and tie each task to a user or customer. You can also reconnect to a task and save its results.Learn from productionTurn real usage into signals, validate the smallest useful change, and distill it back into the recipe.

How the system fits together

An agent recipe runs in production; production experience becomes signals, validated learning, and a new recipe version — Text representation of the diagram:
`Agent recipe → Runtime → Production experience → Signals ↑ │ └──── New recipe version ← Validated learning ┘`

Layer	Responsibility
Agent recipe	The agent, its instructions, skills, tools, and quality criteria in source you own.
Pi	Loads the recipe and runs the agent on your computer or in a managed runtime.
Introspection runtime	Runs one fixed recipe version. It handles customer work in separate sandboxes and saves the results.
Continual learning loop	Collects signals from real use, tests proposed changes, and adds approved changes to the next recipe version.

The model is interchangeable. Your definition of good is not: each validated lesson returns to the agent recipe, where production experience compounds into durable domain IP.

A concrete example

A refund agent promises an exception for a final-sale order. Feedback and observations show the problem, and related conversations show that it has happened more than once. You identify the missing policy decision and add representative test cases. You then test a focused change that fixes the problem without breaking ordinary refund requests. The approved change goes into the recipe as a reviewable Git change.

Choose your next page

If you are new to Introspection, complete the Quickstart.
If you need definitions for product terms, read Core concepts.
If you already run an agent, begin with Learn from production.
If you are building an application, choose an SDK.
If you need a command, use the CLI reference.