Coordinate Multiple AI Agents Through Task Orchestration

Implements a production-tested pattern for coordinating multiple specialized AI agents through a single orchestrator that decomposes tasks, routes them to the right specialist, prevents duplicate or conflicting work, and verifies results before marking anything done - described as battle-tested across 10,000+ tasks over 6 months. The orchestrator is given an explicit identity block stating what it is, and just as importantly a "NOT-block" listing what it must refuse to do (not a code writer, not a researcher, not a tester - always delegate), which the skill reports reduces task drift by roughly 35% in production. Before assigning new work it checks a task registry for near-duplicates using string-similarity matching against pending or in-progress tasks, routes tasks to specialists by keyword-scoring the task description against each agent's keyword list, enforces evidence-based quality gates (confirming files actually changed, tests passing, no secrets introduced, the build succeeding, and only intended files touched) before accepting a "done" claim, and runs a 30-minute heartbeat that reassigns or nudges agents that have gone idle.

Use it when 3 or more specialized agents need to coordinate on complex tasks, when agents are producing duplicate or conflicting work, when an audit trail of who did what and when is needed, or when agent output quality is inconsistent and needs a verification gate before it's trusted. The core discipline is to treat an agent's "done" report as a claim, not evidence, and to mark work complete only after independent checks pass. It is not a substitute for environment-specific validation, testing, or expert review, and its own limitations say to stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.

Input is an incoming task description plus the current task registry (task id, description, assigned agent, status). The pattern checks the description against existing pending/in-progress tasks with a similarity threshold of 0.55 (a lower threshold produces too many false positives) using this check:

import sqlite3
from difflib import SequenceMatcher

def check_duplicate(description, threshold=0.55):
    conn = sqlite3.connect("task_registry.db")
    c = conn.cursor()
    c.execute("SELECT id, description, agent, status FROM tasks WHERE status IN ('pending', 'in_progress')")
    for row in c.fetchall():
        ratio = SequenceMatcher(None, description.lower(), row[1].lower()).ratio()
        if ratio >= threshold:
            return {"id": row[0], "description": row[1], "agent": row[2]}
    return None

Output of a successful delegation is a structured assignment (task, scope, verification command, deadline) logged with task ID, agent, scope, deadline, and verification command, plus the pass/fail result of each quality-gate check.

Uses SQLite as the lightweight task registry (no server required), Python's difflib.SequenceMatcher for duplicate detection, and shell-level verification tooling - git diff --stat, npm test / pytest, npm run build, and secret-pattern grepping - to turn agent completion claims into verified evidence. It's designed to sit alongside related skills for code review, test-driven development, and project-management tracking rather than replace them.

Teams running 3 or more specialized AI agents on the same codebase or project who need a coordination layer to prevent duplicate work, enforce evidence-based completion gates, and keep an audit trail, instead of letting agents self-report progress without verification.