Skip to main content
Agents are best when the task needs code, tools, or per-input adaptation instead of a fixed workflow graph. The source lives in Git, so you work with normal files and commits.

1. Clone the agent source

Use a directory name that makes the organization and environment obvious.

2. Edit the agent

Define the agent in agents/extract-invoice/agent.ts, using the public SDK:
You can move inline instructions into AGENT.md in the same directory. Attach shared resources through typed imports. Validate with eigenpal agents validate agents/extract-invoice; source is compiled in the platform sandbox. Each run captures its compiled definition and exact Git commit in eigenpal.lock. Resumes reuse that snapshot.

3. Commit, push, and release

Run release from the agent package directory (the folder that contains agent.ts), not the organization repo root:
save validates, commits, and pushes local changes. release creates an immutable source tag from the current package, and sync points the automation at the latest release.

4. Run the agent

This example uses structured input and needs no filesystem sandbox. To process uploaded files with shell/read tools, add .withFilesystem() first. The command prints a run id. Poll or watch it like any other run:

5. Improve with reviews

When a run is wrong, add a review and promote the corrected case into the dataset:
The next experiment can now catch the same miss before it ships again. For a full review workflow, see Review production runs. For the full authoring API, see Agent authoring SDK. To start these agents from an application, use the API client.