Skip to content

Runtime

A runtime loads a persona revision, verifies its linked definitions, and resolves the workspace resources needed to run it. It checks permissions and capabilities and stores run state. It uses a harness for work that requires an agent loop.

Resolve linked definitions at their recorded revisions and verify their fingerprints before activation. Studio also resolves the playbook revision pinned by each placement. Capture the definition and input versions for a run so a later import cannot change its meaning while it is active.

Capability describes what the runtime can execute. Setup describes whether this installation has the model, authorized connector account, output destination, or compute needed for a supported step. Report both before dispatch so an operator can distinguish a missing account from an unsupported operation.

A capability report for a Meeting desk installation could present the following information. The labels are illustrative.

Step Runtime capability Workspace readiness
read-meetings supported calendar account bound
prepare-brief supported model bound
publish-brief supported waiting for effect approval
send-sms-reminder unsupported blocked

At dispatch, resolve the acting account and current grants and apply the step’s policy. At an agent step, pass the allowed context, model, and tools to the harness. At a connector write, verify the current request against the reviewed effect before calling the provider.

Restore durable run state and recheck pending authorization. A reconnect or definition sync leaves unsupported steps blocked and completed work complete. If a provider may have received a request before the connection dropped, reconcile the provider outcome before making a second attempt.

See the capability report and runtime rules and execution receipts.