Run

Definition

A Run is one bounded attempt to execute caller-owned Task intent inside a Task Workspace, with an objective, Agent Profile, selected context, execution provider, lifecycle, and result.

Why it exists

Boundaries make agent work observable, repeatable, interruptible, and safe to schedule. A Run can fail or expire without making the whole workspace unusable.

Lifecycle

queued → starting → active → validating → candidate
                       ├→ failed
                       └→ cancelled

Shared or private?

The Run’s status and execution Intent are shared. The Task’s durable intent is owned by its caller; detailed attempt plans and transient scratch state are private by default.

Durable or runtime?

The Run record and selected Events may be durable. Full telemetry, logs, and every intermediate thought do not need permanent storage.

Default implementation

The reference CLI uses an isolated Git worktree, selected repository context, and a command adapter for the agent provider. It can compose from declared System/Layer/Resource membership, but Planner and Task lifecycle ownership remain external.

Trade-offs

Bounded Runs add lifecycle bookkeeping. In return, they make concurrency and repeated scheduled roles composable.