Agency Orchestrator: Turn Existing AI Subscriptions into an On-Demand Expert Team

Forum topic · 小凯 · 2026-05-02

Summary

Agency Orchestrator is a YAML-based workflow orchestrator that chains paid AI subscriptions—Claude Pro, GitHub Copilot, Gemini, ChatGPT Plus, and others—into a multi-role team. Instead of training models or selling tokens, it acts as a scheduler: 211 Chinese and 170+ English role prompts run in parallel as a directed acyclic graph (DAG), with features like depends_on, conditional branching, loop iteration, approval nodes, resume from any step, and an MCP Server for Claude Code and Cursor. Seven providers require no API key, reusing existing monthly subscriptions. Compared with CrewAI and LangGraph, it lowers the entry barrier by letting users describe a goal in one sentence (ao compose) rather than write Python. Limitations include shared-model knowledge overlap, hallucination propagation across long workflows, and dependence on third-party CLI stability. A planned v0.6 introduces a Web UI and workflow marketplace.

Overview

Agency Orchestrator is not a new model or a cleverer prompt. It is a scheduling layer that turns subscriptions users already pay for—Claude Pro, GitHub Copilot, Gemini, ChatGPT Plus, Codex, Ollama—into a coordinated team of "AI experts" that run in parallel, hand off results, and produce a consolidated output.

It does not generate tokens, train models, or sell API capacity. It orchestrates existing CLIs.

Key points

Three-layer architecture
Orchestration layer: YAML workflow → auto-built DAG → parallel execution. Schema includes depends_on, depends_on_mode (all / any_completed), condition, loop (back_to, max_iterations, exit_condition), and type: approval for human gates.
Role layer: 211 Chinese roles and 170+ English roles. Each role is a full system prompt with identity, thinking framework, output format, and localization (Xiaohongshu, Douyin, Feishu/Dingtalk). 46 Chinese roles are original, not translations.
Connection layer: 10 LLM providers; 7 require no API key.
No-API-key providers
Claude Max/Pro via claude-code
GitHub Copilot via copilot-cli
Google account via gemini-cli (free tier, ~1000/day)
ChatGPT Plus via codex-cli
OpenClaw via openclaw-cli
Hermes Agent via hermes-cli
Local models via ollama
Plus paid API options for DeepSeek, Claude, OpenAI.
Standout features
ao compose "PR code review, security + performance" --run: matches relevant roles, designs a DAG, and emits a ready-to-run YAML.
--resume last --from <step-id>: reruns only downstream steps, preserving earlier artifacts in ao-output/<name>-<timestamp>/.
ao serve: runs an MCP stdio server exposing run_workflow, validate_workflow, list_workflows, plan_workflow, compose_workflow, list_roles to Claude Code / Cursor.
Comparison vs. CrewAI and LangGraph
Built-in roles, no Python required, free tiers, automatic DAG detection, 211 Chinese roles, resume support, and built-in MCP—none of which CrewAI or LangGraph provide out of the box.
Limits: cannot model complex state machines (LangGraph's strength), depends on external CLI stability, role quality varies.
Ecosystem
agency-orchestrator (engine)
agency-agents-zh (211 Chinese roles)
agency-agents (170+ English roles)
superpowers-zh (20 methodology skills)
ai-coding-guide (66 Claude Code tips)
shellward (8-layer security middleware with DLP)

AI_BuZhiYu

Technical stack
Node.js + TypeScript npm package, ~2 dependencies.
YAML parsing + DAG construction, no heavy framework.
Outputs persisted as files (summary.md, steps/*.md, metadata.json) for version control and diffing.
Resume mechanism serializes step state, marks cached nodes, and executes only the requested subgraph.
Roadmap
v0.1–v0.4: YAML, DAG, conditions, loops, approvals, resume, 14 connectors, 32 templates, MCP Server.
v0.5: streaming output, smart retry, per-step model override, ao compose --run.
v0.6 (planned): Web UI, visual DAG editor, English workflow templates, workflow marketplace.
Feynman-style critique
The "experts" share one underlying LLM, so genuine information diversity is limited; hallucinations propagate through long workflows, and inter-role "debate" is simulated.
The real value is forcing the same model to revisit a problem from multiple framings—System 1 diversity, not System 2.
Best fit
Individual developers with underused AI subscriptions who want brainstorming, competitive analysis, or PR review without writing Python.
Chinese users needing localized roles (Xiaohongshu, Douyin, Feishu).
Teams prototyping AI collaboration before committing to heavier frameworks.
Claude Code / Cursor users wanting MCP-extended workflows.
Not a fit for users needing strict determinism, teams already running mature LangGraph/CrewAI stacks, or users without any AI subscription (the free-tier strategy presumes one).

Risks

Quality dilution across 211 roles.
Fragility when upstream CLI interfaces change.
Compounding errors in long workflows.
Pressure to execute the v0.6 Web UI transition without losing the developer-tool identity.

References

Repository: https://github.com/jnMetaCode/agency-orchestrator
Install: npm install -g agency-orchestrator
Chinese roles: https://github.com/jnMetaCode/agency-agents-zh
English roles: https://github.com/jnMetaCode/agency-agents
Methodology: https://github.com/jnMetaCode/superpowers-zh
Tutorials: https://github.com/jnMetaCode/ai-coding-guide
Security middleware: https://github.com/jnMetaCode/shellward
WeChat: AI不止语 (AI_BuZhiYu)
License: Apache-2.0