Documentation/FoundationsDownload prototype ↓

FOUNDATIONS / DESIGN PARTNER PREVIEW

A capability has a lifecycle.

Code, data, authority, and versions belong to the same unit of useful software.

Capability

The useful thing an agent can run. In the prototype, a capability is a stored JavaScript module with a tenant, an identifier, a version and an enabled flag. The managed product will need a richer definition of inputs, permitted operations, resource ownership and retention.

Trusted host

The host authenticates the caller, selects the tenant, loads the tool, and decides what to pass into its execution environment. Generated code does not decide its own authority. The prototype passes a snapshot of the authenticated tenant’s records through env.INPUT.

Isolated execution

A Dynamic Worker runs the stored module. The prototype does not give it a database handle or a cloud credential. Outbound networking is disabled. This limits the particular operations a tool can perform; it is not a complete sandbox security audit.

Persistent data versus infrastructure state

Application data is the information a tool uses—research records, for example. Infrastructure state records which cloud resources belong to a deployment and what reconciliation should do next. Keeping application data is not the same as making deployments recoverable.

Versions and revocation

The prototype uses optimistic version checks when changing records or tools. A stale update returns a conflict. Revoking a tool prevents subsequent execution requests. It does not cancel already-running work or erase stored records.

The lifecycle

  1. Define: establish the tool, resources, and allowed inputs.
  2. Run: execute inside the granted boundary.
  3. Keep: preserve useful code and data between sessions.
  4. Evolve: version behavior while retaining identity.
  5. Retire: revoke execution and explicitly decide what data to retain.