Skip to content

Agent Runtime first workflow scope

The first workflow slice is intentionally bounded. It composes existing capabilities and registered tools; it does not create a general-purpose scheduler or autonomous agent loop.

The initial coordinator implementation executes one declared workflow of sequential tool steps and explicitly declared parallel groups with the limits below. Persistence and autonomous replanning remain outside this first slice.

Sequential steps and bounded parallel groups are now implemented by the workflow coordinator; the limits below remain enforced at the contract boundary. Admission validation rejects duplicate step IDs, empty or oversized parallel groups, and workflows exceeding sixteen total child steps before any handler starts.

Allowed execution model

  • One parent workflow per request.
  • Sequential child steps and explicitly declared parallel groups.
  • Maximum depth of 4 nested workflows.
  • Maximum of 16 child steps per workflow and 8 concurrent children per group.
  • A finite workflow deadline inherited from the request and bounded by runtime configuration.

Each child must be declared before execution, use an existing capability or registered tool, and inherit the parent execution context. Dynamic discovery may select from sealed registries but may not create hidden handlers.

Side-effect boundaries

The workflow may pass typed, bounded results between declared steps. It may not execute arbitrary code, access the filesystem or network implicitly, inject credentials, or persist state unless an explicitly registered tool owns that side effect and policy authorizes it.

Retries, fallback, and provider routing remain governed by their existing contracts; the workflow does not multiply those policies autonomously.

Admission and termination

Admission fails before child execution when depth, fan-out, deadline, or policy limits are exceeded. Caller cancellation and deadline timeout stop admission, cancel active children cooperatively, and preserve the workflow result and error precedence contracts. Cleanup runs in reverse declaration order.

Non-goals

  • unbounded recursion or fan-out;
  • autonomous replanning loops;
  • implicit retries or provider selection;
  • durable workflow storage;
  • multi-agent communication or memory; and
  • a hosted scheduler, approval service, or policy engine.