Coordinator

Coordinator Pattern

The coordinator is the orchestrator of multi-agent workflows. It plans work, delegates to specialists, monitors progress, and reports results. It does not write code — it coordinates the agents that do.

Overview

A coordinator agent (type koordinator) sits between you and the worker agents. You give it a high-level directive, and it handles the rest:

  • Plan — read the codebase, understand the task, break it into agent-sized pieces
  • Delegate — spawn the right agent types with specific tasks
  • Monitor — check agent output, course-correct when agents go off track
  • Report — synthesize results and deliver a clean summary

Delegation, Not Implementation

The most important rule: coordinators do not write code. If a task requires editing files, implementing features, or debugging, that's a worker agent's job. The coordinator spawns the right agent type and gives it clear instructions.

Coordinator DoesWorker Agents Do
Plan the work splitRead and write code
Assign tasks to agentsInvestigate codebases
Monitor progressFix bugs and implement features
Course-correct when stuckWrite tests and documentation
Report results to youCommit and verify changes

Monitoring Agents

The coordinator needs visibility into what agents are doing. These tools provide that:

ToolDescription
hub_agentsList all online agents with status
hub_statusFull roster + coordinator info
hub_captureView an agent's recent output
hub_workView work queue and active assignments
claimsList all active file lane claims
# Check what agents are doing
<hub_capture name="sapphire" lines="100"/>
# See who's online
<hub_agents/>
# View work queue
<hub_work/>

Work Assignment Patterns

There are two main patterns for getting work to agents:

Pattern 1: Spawn with Task

Spawn an agent with a task description. The agent starts working immediately.

# Spawn with a specific task
/hub spawn coder fix the auth bug in login.py. check tests pass.

Pattern 2: Spawn then Message

Spawn an agent first, then send follow-up instructions via hub_msg. Useful when the task evolves.

# Spawn the agent
/hub spawn coder
# Send initial task
/hub msg sapphire refactor the auth module. start with login.py
# Send follow-up
/hub msg sapphire also check the session handler - it might need the same treatment

When to Use a Coordinator

Not every task needs a coordinator. Here's when it makes sense:

Use CoordinatorUse Single Agent
Multi-file changes spanning different areasSingle file or focused bug fix
Need parallel execution (frontend + backend + tests)Sequential work is fine
Complex task that benefits from specialist agentsSimple task any agent can handle
Want progress monitoring and course correctionFire and forget is acceptable
Need a clean summary of what happenedYou can review the output directly

Coordinator Lifecycle

A typical coordinator workflow follows this sequence:

1. Acknowledge — tell you "on it" immediately
2. Investigate — read relevant files, understand the scope
3. Plan — break work into 2-4 agent tasks
4. Spawn — create agents with specific assignments
5. Monitor — check output, course-correct as needed
6. Verify — review completed work before reporting
7. Report — deliver a clean summary of results

The coordinator stays available throughout — you can ask questions, change direction, or add tasks at any point.

Next Steps