Choose your path
Agent Foundation supports managed execution through Service, embedded execution through Harness, and interactive work through Harness UI.
| Your goal | What you need | Start here |
|---|---|---|
| Deploy a shared agent platform | Docker Compose and a model provider account for the local trial | Service quickstart |
| Use your team's agents | Console URL, an account, and permission to run an agent | Use an existing platform |
| Call remote agents from an application | Service URL, workspace API key, and a configured agent | Connect your application |
| Run agents inside your Python process | Python 3.13+ and uv; the first example uses an offline model | Harness quickstart |
| Work interactively in a repository | Harness UI and a supported model subscription or API key | Harness UI setup |
Choose where execution lives
- Service: your application submits work to a running deployment. Service owns identities, saved conversations, and durable execution; Console is its browser application.
- Harness: your Python process constructs and runs the agent. Your application owns storage, access policy, and delivery.
- Harness UI: the terminal and browser workbench runs agents with its own configuration and conversation history. Share an instance with trusted collaborators.
A Service SDK is a client for remote execution. The Harness library executes agents inside your process. Service Console and Harness UI's browser are separate applications.
Use individual components
| Need | Guide |
|---|---|
| Portable files and commands, with or without an agent | Environments |
| Environment operations through a daemon | Envd |
| AG-UI events from Harness observations | Stream Protocol |
| Distribution names, imports, and runnable examples | Package catalog |
For a guided explanation of one conversation, read Core concepts. For deployment operations, see Run and maintain. Service clients have independent versions; check their supported contract against your deployment.