Skip to main content
An agent product combines two kinds of machinery:
  1. The Harness runs the agent loop: it prepares model context, calls a model, invokes tools, and decides what to do next.
  2. The platform is everything that has to outlive that loop: durable state, governance, and a record of what happened.
Checkfu owns the second layer and treats the first as replaceable.

The boundary

Your product calls Checkfu’s HTTP surface. Checkfu resolves the Agent’s published definition, provisions a sandbox, and launches its configured Harness. Model calls, capabilities, credentials, and Memory cross platform-owned boundaries so policy and accounting remain consistent across harnesses.

Capabilities are searched; data is mounted

Checkfu keeps capability descriptions out of the model context until they are needed. Tools, skills, and Memory recall live in searchable catalogs. Data such as repositories, files, and Memory directories mounts into the sandbox filesystem, where it does not consume model context merely by existing. So a Workspace can carry a large catalog and a large repository without either one filling the model’s context window before the first turn.

Ports, not permanent dependencies

Each infrastructure plane sits behind a replaceable port. Local development can use in-memory stores, a local sandbox, and a mock model; production can use hosted adapters or customer-side infrastructure. The same domain behavior is expected from each adapter.

Next steps

Sessions and Runs

See how the control plane turns an Agent launch into an event-sourced Session and its sandboxed Runs.

API reference

Explore the HTTP surface your product calls to resolve, launch, and observe agents.