managed-team

name	managed-team
description	Use when facing multiple tasks that need parallel execution by a team of named agents pulling from a differ-review kanban board, or when asked to find and fix bugs across a codebase with coordinated workers

Managed Team

Spin up a team of named agents that self-direct from a differ-review kanban board. You act as project manager.

Modes

Execute Board

Tasks already exist on the board. Spawn workers, they pull and complete tasks.

/managed-team execute

Discover and Fix

No tasks yet. Spawn discovery agents (pentesters, quality reviewers), create tasks from findings, then spawn fixers.

/managed-team discover-and-fix

Core Pattern

digraph managed_team {
    "Determine mode" [shape=diamond];
    "Board has tasks?" [shape=diamond];
    "Spawn discovery agents" [shape=box];
    "Create tasks from findings" [shape=box];
    "Generate names" [shape=box];
    "Spawn N workers" [shape=box];
    "Monitor board" [shape=box];
    "All tasks done?" [shape=diamond];
    "Run verification" [shape=box];
    "Report" [shape=box];

    "Determine mode" -> "Board has tasks?" [label="execute"];
    "Determine mode" -> "Spawn discovery agents" [label="discover"];
    "Spawn discovery agents" -> "Create tasks from findings";
    "Create tasks from findings" -> "Generate names";
    "Board has tasks?" -> "Generate names" [label="yes"];
    "Board has tasks?" -> "Spawn discovery agents" [label="no"];
    "Generate names" -> "Spawn N workers";
    "Spawn N workers" -> "Monitor board";
    "Monitor board" -> "All tasks done?";
    "All tasks done?" -> "Monitor board" [label="no"];
    "All tasks done?" -> "Run verification" [label="yes"];
    "Run verification" -> "Report";
}

Setup

1. Generate Names

mcp__differ-review__random_name()  # one per worker

2. Worker Prompt Template

Each worker gets this prompt (fill in name and constraints):

You are {NAME}, a fixer agent. Pick up tasks from the differ-review kanban board and fix them.

WORKFLOW:
1. mcp__differ-review__list_tasks(repo_path="{REPO}") to see pending tasks
2. Pick highest-priority pending task
3. mcp__differ-review__take_task(task_id=..., worker_name="{NAME}") to claim it
4. Do the work (read, fix, test)
5. mcp__differ-review__update_task(task_id=..., status="done", note="summary")
6. Go back to step 1

CONSTRAINTS:
{CONSTRAINTS}

3. Spawn Workers

# 3-4 workers is the sweet spot. More causes contention.
Task(name=name, subagent_type="general-purpose", mode="bypassPermissions",
     team_name=team, prompt=worker_prompt)

Project Manager Responsibilities

Monitor

Check mcp__differ-review__list_tasks periodically to track progress.

Create Tasks from Findings

When discovery agents report, create tasks with severity in title:

mcp__differ-review__create_task(repo_path=REPO, title="[HIGH] Fix X", description="...")

Handle Stale Agents

If an agent goes idle without completing its task, send a nudge via SendMessage. If unresponsive after 2 nudges, shut down and respawn.

Avoid Full Test Suite

Broadcast to workers: run only targeted tests relevant to their change. Full suite runs should be a separate task at the end.

Discovery Agent Types

For discover-and-fix mode, spawn these as background agents:

Type	Focus	Count
Pentester	Auth, access control, crypto, input validation	2-3
Quality	Code bugs, UX issues, performance	1-2

Discovery agents report findings but do NOT fix. Their output becomes tasks.

Constraints to Always Include

- Use {VENV_PATH} for all Python/pytest commands (if applicable)
- Do NOT use git stash
- Run only targeted tests, not the full suite
- Do NOT modify files outside your task scope

Common Mistakes

Mistake	Fix
Too many workers (5+)	Stick to 3-4. More causes file contention.
Workers run full test suite	Broadcast: targeted tests only.
Stale shutdown requests kill respawned agents	Use fresh names for respawns.
Workers duplicate each other's tasks	differ-review take_task prevents this.
No verification at end	Always run full test suite as final task.