Delegation & Human Tasks

Delegation & Human Tasks

Preview

The durable delegation foundation is passing focused tests for child success and failure, but the full tracked Agent Runtime delegation flow is still deliberately disabled.

Delegation is easy when it is just “ask another model and wait.”

It gets harder when the parent task must survive restarts, wait on a child, handle failure, verify the result, or stop and ask a person for help.

LM Nexus is building those cases into durable work instead of treating delegation as a nested chat call.

Human Tasks

A Human Task represents work that cannot continue until a person provides input or performs an action.

Runtime Supervisor can turn a blocked execution into a Human Task instead of leaving it silently stuck.

That gives durable work an explicit state for “automation needs a person now.”

Durable child Tasks

The current implementation can create durable child Tasks and feed their final results back into the parent work.

Focused tests cover both:

• verified child-success completion/resume foundations;
• terminal child failure feeding the parent dependency-failure / Supervisor path.

Why full tracked delegation is not enabled yet

Two important boundaries still need to land.

1. Work Verification / Acceptance

A parent needs a reliable, replay-safe way to decide whether a child’s output should be accepted.

The intended contract uses explicit outcomes such as PASS, FAIL, and INCONCLUSIVE and records where that verification decision came from.

2. Parked-parent coordination

A synchronous parent Agent Runtime loop should not keep running while its durable parent Task is waiting for a child.

The runtime needs a durable way to park, wait, and resume the parent instead of pretending “delegated” and “completed” happen at the same moment.

Current delegation versus durable tracked delegation

Agent Runtime already supports limited delegation today.

The Preview work described here goes further: durable child Tasks, restart-aware completion and failure, reliable verification, parked parents, and replay-safe resume.

They are related features, but they are not at the same maturity level yet.