| name | delegation-orchestration |
| description | Coordinate Sloppy child agents through the delegation provider without polling, including parent-side parallel work, result retrieval, and child-session cleanup. |
| version | 1.0.0 |
| metadata | {"sloppy":{"tags":["runtime","delegation","orchestration"],"category":"runtime"}} |
Delegation Orchestration
When To Use
Use this skill when a task asks for child agents, sub-agents, parallel exploration, delegated review, or merging delegated findings.
Procedure
- Spawn the requested child agents with focused goals and clear read/write boundaries.
- If the user asked you to also work "in the meantime" or "in parallel", do your own independent work before the first delegation wait.
- Join child progress with
slop_wait_for_delegation_event; do not repeatedly query /delegation/agents as a polling loop.
- Treat each wait as one wake event. Wait again when more children remain active.
- Call
get_result for each completed child before relying on its findings.
- If a child needs a follow-up, use
send_message and wait again.
- After retrieving a child's final result, call
close unless you still need a follow-up turn.
- Merge child findings with your own findings and identify which conclusions came from which source when useful.
Pitfalls
- Do not treat
spawn_agent success as a completed result.
- Do not skip parent-side work when the user explicitly requested it.
- Do not leave completed child sessions open after the final result is retrieved.
- Do not synthesize from
result_preview alone when get_result is available.
Verification
Before final response, confirm every spawned child is either still intentionally open for follow-up or has been closed after get_result.
