Use when an org role acts as queen coordinator and must decompose, delegate and reconcile work within one bounded session whose workers report only to it. Covers authoritative session state, capability routing, acceptance gates, drift correction and not overclaiming consensus. For a multi-level manager/worker tree, use hierarchical-coord.
Installs into .claude/skills of the current project.
Are you the author of Queen Coordinator?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/monoes-queen-coordinator)
---
name: queen-coordinator
description: "Use when an org role acts as queen coordinator and must decompose, delegate and reconcile work within one bounded session whose workers report only to it. Covers authoritative session state, capability routing, acceptance gates, drift correction and not overclaiming consensus. For a multi-level manager/worker tree, use hierarchical-coord."
tags: ["operations","coordination","agents"]
tools: []
license: Apache-2.0
source: https://github.com/monoes/monomind
---
# Queen Coordinator — Best Practices
## Focus
Same shape as a lead coordinator — decompose, delegate, hold authoritative state, decide when done — but scoped to a single bounded session with workers that report directly up to it and nowhere else.
## Best practices
- Own the single source of truth for the session: what's in-progress, blocked, done, and reconciled. Worker reports are inputs to that state, not the state itself.
- Decompose the session's objective into subtasks with one accountable worker and explicit acceptance criteria each before dispatching anything.
- Route by capability — match each subtask to the worker actually suited for it rather than whichever worker is idle.
- Apply the session's approval policy before accepting a deliverable; block anything that fails acceptance criteria with specific, actionable feedback rather than accepting to keep moving.
- Detect drift early and re-scope immediately — a worker diverging from the session goal costs more the longer it's left uncorrected.
- Close the session explicitly: reconcile all final worker state into one authoritative record before declaring done, not just when the last worker happens to report in.
## Common pitfalls
- Treating "queen" as implying capabilities the session doesn't actually have — no background timers, no session succession, no automatic redistribution based on inferred worker load. If something must outlive the session, persist it before returning; nothing carries it forward automatically.
- Rubber-stamping the last worker report instead of reconciling all reports into consistent authoritative state.
- Letting two workers silently take overlapping scope because delegation wasn't explicit about boundaries.
- Describing the session as achieving "consensus" or being "fault tolerant" when it was actually one coordinator making decisions — nothing tolerated a fault, no vote was tallied.
- Scoping the session too broadly, so "queen" coordination is really trying to run a whole hierarchical org through one session instead of one bounded unit of work.
## Tools & techniques
- Dispatch independent worker tasks in a single batch — that's the only source of real concurrency; the queen doesn't grant parallelism by existing, the dispatch mechanism does.
- Keep session state (roster, task assignments, acceptance status) in one inspectable place so reconciliation is a read, not a reconstruction.
- Use an explicit vote/threshold mechanism (not the queen's own judgment) for any decision that genuinely needs multiple workers to agree — don't conflate "queen decides" with "workers voted."
- Route synthesis-across-workers to a knowledge-consolidation step distinct from task routing — deciding who does what and reconciling what they collectively learned are different jobs, even in one session.