Coordinate Multiple AI Agents Through Task Orchestration

Multi-Agent Task Orchestrator is a pattern for coordinating several specialized AI agents through a single orchestrator instead of letting them work independently and conflict, described as battle-tested across 10,000-plus tasks over 6 months. The orchestrator is given an explicit identity block stating what it is NOT - not a code writer, not a researcher, not a tester - which reduces task drift by about 35% in production by stopping it from doing specialist work instead of delegating. Before assigning anything, it checks a SQLite task registry for near-duplicate work using a string-similarity ratio (SequenceMatcher) at a 55% threshold, routes new tasks to a specialist agent by keyword-scoring the description against each agent's keyword list, and enforces evidence-based quality gates - an agent's completion report is a claim, not evidence - before marking anything done:

After agent reports completion:
1. Were files actually modified? (git diff --stat)
2. Do tests pass? (npm test / pytest)
3. Were secrets introduced? (grep for API keys, tokens)
4. Did the build succeed? (npm run build)
5. Were only intended files touched? (scope check)

Mark done ONLY after ALL checks pass.

Every 30 minutes it also runs a heartbeat: check what it delegated in that window, pull the next task from backlog if nothing was delegated, and reassign any agent that has gone quiet for over 30 minutes on its assigned task.

Use it once there are three or more specialized agents that need to coordinate on complex work, when agents are duplicating or conflicting on the same task, 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 acceptance. Named agent roles in the reference implementation are keyword-routed: code-architect (code, implement, function, bug, fix, refactor, api), security-reviewer (security, vulnerability, audit, cve, injection), researcher (research, compare, analyze, benchmark, evaluate), doc-writer (document, readme, explain, tutorial, guide), and test-engineer (test, coverage, unittest, pytest, spec).

Input is an incoming task description. Output is a routed, deduplicated assignment - to the highest-scoring specialist, or a skip decision with the existing task ID if a similar one is already in progress - plus a completion verdict that only passes once git diff, test runs, a secret scan, and a build check all clear, and a delegation log recording task ID, agent, scope, deadline, and verification command for every assignment.

SQLite for the task registry (lightweight, no server needed); git diff --stat, test runners (npm test, pytest), a build command (npm run build), and a secrets grep as the quality-gate checks. It pairs with @code-review for reviewing delegated code changes, @test-driven-development for output quality, and @project-management for tracking overall progress.

Anyone running three or more specialized agents on the same codebase or project who is seeing duplicate work, silent scope creep, or agents marking tasks done without real evidence that the work happened.