agent-coordinator

Agent Coordinator

Coordinate multiple AI coding agents — architect, developer, QA — on the same codebase. Agents pass work through handoff.md; the coordinator reads who's next, builds the prompt, dispatches to the backend, verifies the handoff, and repeats.

Works with GitHub Copilot CLI, Claude Code, OpenCode, or any CLI tool. Mix backends freely. Zero third-party dependencies.

Shared files:

• handoff.md — agent conversation history, append-only
• tasks.json — task state, updated automatically
• .agent-coordinator/ — coordinator state (sessions, event log, debug log)

Install

pip install agent-coordinator

uv run pipx install agent-coordinator

From source:

git clone https://github.com/zkucekovic/agent-coordinator.git
cd agent-coordinator
pip install -e .

Requires Python 3.10+. Install whichever backend CLI you plan to use.

Quick start

Option A — Start from scratch

# Initialise a new workspace (creates handoff.md + agents.json automatically)
agent-coordinator --workspace ./my-project

The coordinator creates two files in ./my-project/:

agents.json — ready to run, no editing needed:

{
  "default_backend": "copilot",
  "retry_policy": { "max_rework": 3, "on_exceed": "needs_human" },
  "agents": {
    "architect":   { "backend": "copilot", "model": "claude-sonnet-4.6", "prompt_file": "prompts/architect.md", "supportsStatelessMode": false },
    "developer":   { "backend": "copilot", "model": "claude-sonnet-4.6", "prompt_file": "prompts/developer.md", "supportsStatelessMode": true },
    "qa_engineer": { "backend": "copilot", "model": "claude-sonnet-4.6", "prompt_file": "prompts/qa_engineer.md", "supportsStatelessMode": true }
  }
}

handoff.md — initial handoff directing the architect to start planning.

Then drop in your project context and hit Start:

# Add a spec and/or plan (optional — agents will work from handoff.md otherwise)
cp SPECIFICATION.md ./my-project/
cp plan.md ./my-project/

# Run — the startup menu appears; press Enter to start
agent-coordinator --workspace ./my-project

---

Option B — Import specs and plans, then run

Import your existing documents and let the architect pick up immediately:

# Import a single spec or plan (auto-detected)
agent-coordinator --import SPECIFICATION.md --workspace ./my-project
agent-coordinator --import plan.md --workspace ./my-project

# Import entire folders
agent-coordinator --import-specs ./specs/ --workspace ./my-project
agent-coordinator --import-plans ./plans/ --workspace ./my-project

# Then start the coordinator
agent-coordinator --workspace ./my-project

On first run the coordinator injects all spec/plan files into the architect's prompt. The architect reads them, decomposes the work into tasks.json, and hands off task-001 to the developer.

---

The development loop

Once running, each turn follows this cycle:

architect  →  reads spec/plan, creates tasks.json, assigns task-001 to developer
developer  →  implements the task, writes code, updates handoff.md (STATUS: review_required)
architect  →  reviews the work (STATUS: approved or rework_required)
qa_engineer → validates, runs tests, writes handoff.md (STATUS: continue or blocked)
              ↑ repeat for each task

Working through a task list:

The coordinator tracks tasks.json automatically. After each approved task the architect assigns the next one. You can watch progress live in the TUI or check tasks.json:

{ "id": "task-001", "title": "...", "status": "done" },
{ "id": "task-002", "title": "...", "status": "in_engineering" },
{ "id": "task-003", "title": "...", "status": "planned" }

To skip to a specific task or add context, press Ctrl+C → m (add message to handoff) and type your instruction.

---

Skip the menu — run immediately

agent-coordinator --workspace ./my-project --auto          # skip startup menu
agent-coordinator --workspace ./my-project --max-turns 20  # stop after 20 turns
agent-coordinator --workspace ./my-project --reset         # clear sessions and restart

How it works

---HANDOFF---
ROLE: developer
STATUS: review_required
NEXT: architect
TASK_ID: task-001
TITLE: Implement login endpoint
SUMMARY: Added POST /auth/login with JWT signing. Used existing User model.
ACCEPTANCE:
- login endpoint returns a signed JWT — PASS
- input validation on email/password — PASS
CHANGED_FILES:
- src/auth/login.py
- tests/test_login.py
VALIDATION:
- python3 -m pytest tests/test_login.py -- 6 passed
BLOCKERS:
- none
---END---

Each turn: read NEXT: → build prompt → dispatch → verify new block appended (retry if not) → sync tasks.json → log to .agent-coordinator/events.jsonl.

Status values

Task lifecycle

planned → in_engineering → ready_for_architect_review → done
                ↑                      ↓
          rework_requested ←── (architect decides)

Configuration

agents.json

{
  "default_backend": "copilot",
  "retry_policy": { "max_rework": 3, "on_exceed": "needs_human" },
  "agents": {
    "architect": {
      "backend": "claude",
      "model": "claude-sonnet-4",
      "prompt_file": "prompts/architect.md",
      "supportsStatelessMode": false
    },
    "developer": {
      "backend": "copilot",
      "prompt_file": "prompts/developer.md",
      "supportsStatelessMode": true
    },
    "qa_engineer": {
      "backend": "opencode",
      "prompt_file": "prompts/qa_engineer.md",
      "supportsStatelessMode": true
    }
  }
}

Each agent can use a different backend. Built-in: copilot, claude, opencode, manual (human-in-the-loop). Any other value is resolved from PATH, or use backend_config for full control — see docs/custom-backends.md. Set supportsStatelessMode: false for agents that should keep their session even when the CLI runs with --stateless.

CLI flags

Project files

Drop these in your workspace — they're injected into agent prompts on the first turn:

• Spec folder: specs/, spec/, or specifications/ — all .md files loaded
• Plan folder: plans/, plan/, or implementation_plans/ — all .md files loaded
• Single spec file (fallback): SPECIFICATION.md, spec.md, PRD.md, requirements.md
• Single plan file (fallback): IMPLEMENTATION_PLAN.md, plan.md
• Project rules: AGENTS.md — coding standards, conventions, and constraints

Folders take priority over single files. Use --import-specs / --import-plans to populate these from outside the workspace.

Sessions

Session IDs are saved in <workspace>/.agent-coordinator/sessions.json. Re-running resumes where you left off. Use --reset to start clean. With --stateless, only agents whose config allows stateless mode get a fresh context every turn; agents with supportsStatelessMode: false keep using their saved sessions.

Interactive control

Press Ctrl+C during any turn:

  c - Continue execution
  r - Retry current turn
  e - Edit handoff.md in $EDITOR
  m - Add message to handoff
  i - Inspect current state
  q - Quit

The workflow also pauses automatically on NEXT: human or when max_rework is exceeded.

Adding agents

No code changes needed. Create a prompt file, add the agent to agents.json, and route to it by writing NEXT: <role> in a handoff block.

"security_reviewer": {
  "backend": "claude",
  "prompt_file": "prompts/security_reviewer.md"
}

Adding backends

For CLI tools, use backend_config in agents.json — no Python needed:

"developer": {
  "backend": "custom",
  "backend_config": {
    "command": ["my-cli", "run"],
    "message_arg": "{message}",
    "workspace_arg": ["--dir", "{workspace}"],
    "session_arg": ["--session", "{session_id}"],
    "output_format": "json",
    "json_text_field": "result"
  }
}

Or implement the AgentRunner interface in Python:

from agent_coordinator.application.runner import AgentRunner
from agent_coordinator.domain.models import RunResult

class MyRunner(AgentRunner):
    def run(self, message, workspace, session_id=None, model=None, on_output=None):
        response = call_my_tool(message)
        return RunResult(session_id="some-id", text=response)

Register in agent_coordinator/cli.py and reference as "backend": "my_runner" in agents.json.

Retry behavior

If an agent doesn't update handoff.md, the coordinator retries with a reminder. If max_rework is exceeded, it escalates per the on_exceed policy.

Demo

agent-coordinator --workspace examples/tetris-demo --max-turns 30

Builds a playable HTML Tetris game through the full architect → developer → QA loop. See examples/tetris-demo/.

Tests

python3 -m unittest discover tests/ -v                                      # unit tests
python3 -m unittest tests.test_handoff_parser -v                            # single file
RUN_INTEGRATION_TESTS=1 python3 -m unittest discover tests/integration/ -v  # requires API tokens

Project structure

agent_coordinator/
  cli.py                  entry point and orchestration loop
  domain/                 models, task lifecycle, retry policy (no I/O)
  application/            router, prompt builder, task service, runner interface
  infrastructure/         backend runners, PTY subprocess, TUI, file I/O
  handoff_parser.py       regex parser for handoff blocks
  prompts/                role instructions and shared protocol rules
  helpers/                import/export utilities
tests/                    unit and integration tests
docs/                     protocol spec, workflow details, backend guide
examples/                 tetris demo, sample configs

Further reading

• ARCHITECTURE.md — hexagonal design, data flow, extension points
• docs/protocol.md — handoff block specification
• docs/workflow.md — coordinator loop and task lifecycle
• docs/custom-backends.md — any CLI as a backend
• docs/interactive-control.md — Ctrl+C menu and human intervention

License

MIT