agents-orchestrator

Trigger phrases

Phrases that activate this skill when typed to Claude Code:

• orchestrate the pipeline
• run the full workflow
• spawn agents-orchestrator

What it does

agents-orchestrator is the conductor persona in the agency-agents collection. It takes a single project spec, then drives the whole pipeline: spawn project-manager-senior to extract a task list, spawn ArchitectUX to lay technical foundations, then enter a per-task loop that alternates a developer agent with an EvidenceQA agent until QA returns PASS — and only then advances to the next task. After all tasks pass, it spawns testing-reality-checker for final integration sign-off.

The discipline the persona enforces is the gate: no task advances on a soft pass, retries are capped at 3, and a fourth failure escalates rather than papering over the issue. It’s the opposite of “one prompt, hope for the best” — it’s a finite-state machine described as a system prompt.

When to use it

• A project spec exists and you want an autonomous run rather than steering each phase yourself
• The work decomposes cleanly into tasks where each one has a verifiable PASS/FAIL outcome
• You want explicit retry-and-escalate behavior instead of silent partial completion
• You’re already comfortable spawning multiple subagents and want a coordinator on top of them

When not to reach for it:

• Exploratory work where the spec doesn’t exist yet — the orchestrator expects a project-specs/*-setup.md file as input
• Single-task changes — the loop overhead isn’t worth it for a one-file fix
• Work that can’t produce screenshot or evidence-based QA — the gate becomes vibes, which defeats the purpose

Install

From msitarzewski/agency-agents at specialized/agents-orchestrator.md. Copy the file to ~/.claude/agents/ (or use the repo’s scripts/install.sh). The orchestrator references a roster of other personas from the same repo (project-manager-senior, ArchitectUX, EvidenceQA, testing-reality-checker, plus per-task developer agents), so it works best when the rest of the agency-agents collection is also installed.

What a session looks like

1. Phase 1 — PM. Orchestrator reads project-specs/<project>-setup.md, spawns project-manager-senior, waits for project-tasks/<project>-tasklist.md to land.
2. Phase 2 — Architecture. Spawns ArchitectUX to produce CSS scaffolding and project-docs/<project>-architecture.md from the spec + task list.
3. Phase 3 — Dev↔QA loop. For each task:
   • Spawn the right developer agent (Frontend Developer, Backend Architect, Mobile App Builder, etc.) scoped to that one task.
   • Spawn EvidenceQA to validate, requiring screenshot evidence.
   • On PASS: advance, reset retry counter. On FAIL: loop back to dev with the QA feedback. After 3 fails: escalate, mark task blocked, continue.
4. Phase 4 — Integration. Once all tasks pass, spawn testing-reality-checker for end-to-end validation. The reality-checker defaults to NEEDS WORK unless evidence proves otherwise.
5. Status reporting. The orchestrator emits structured progress reports between phases — current phase, task completion ratio, retry counts, last QA feedback.

The discipline that makes it work: the per-task QA gate. Without it, the orchestrator collapses into “spawn a bunch of agents and pray” — which is what most multi-agent prompts already do.

Receipts

2026-05-05 — Coordinating six parallel persona dispatches for a receipts batch

Used as the conductor pattern across a 9-skill receipts batch: 3 done by main agent, 6 dispatched to personas in parallel — code-reviewer on a recent commit, evidence-collector on a live page pair, technical-writer on the receipts template, accessibility-auditor on the homepage, codebase-onboarding-engineer on this repo, git-workflow-master on current branch state.

All six were scoped to real artifacts (a committed SHA, two live URLs, a gitignored template, the actual repo state) and all six returned in one foreground response. Total agent tokens ~390K; slowest dispatch was 691 sec (accessibility-auditor with extensive WebFetch). One actionable finding per dispatch landed: a missing frontmatter field on 4 Osmani skills (code-reviewer), a discoverability gap on the SEO topic page (evidence-collector), 5 specific template edits (technical-writer), 3 WCAG AA gaps (accessibility-auditor), a shippable onboarding doc (codebase-onboarding-engineer), and a coherent commit recommendation (git-workflow-master).

Honest scope note: none of these dispatches ran the actual persona file from ~/.claude/agents/. They ran general-purpose agents told “act with X persona’s framing.” The findings are real; the persona-specific behavior is approximated. The full PM → Architect → Dev↔QA-loop → integration pipeline the orchestrator is designed for didn’t run here — this was the conductor pattern applied to a flat fan-out, not the gated multi-phase shape.

Source and attribution

From Michael Sitarzewski’s agency-agents repository, an MIT-licensed collection of 144+ AI agent personas across engineering, marketing, design, testing, and specialized roles.

License: MIT.

Quote from the persona body, verbatim: “No phase advancement without meeting quality standards.” The whole orchestrator is built around making that gate stick.