GitHub - drivelineresearch/moa-x: Mixture-of-Agents-X - utilizing the MoA framework in agentic coding tools.

Cross-Lab Mixture of Agents for coding plans.

A small, CLI-native take on the 2024 Mixture-of-Agents paper, pointed at a different job: producing repo-grounded implementation plans for coding agents instead of chat answers. Three proposers from three different labs (OpenAI codex, Google gemini, Anthropic claude Sonnet) read the repo in parallel, do their own web research, and each write an independent plan. Two of them then refine in broadcast mode (every refiner sees every plan). Finally a parent Claude Opus session aggregates the whole thing into one plan you can act on.

Built to run inside Claude Code as a skill. Standalone Python works too. The harness currently wires up the three vendor CLIs on their subscription plans, which is what I personally run. API-based auth, alternative harnesses, and other model providers are all fair game. See "PRs we'd love to see" below.

TL;DR

# 1. Install the three CLIs (see docs/install.md for details)
npm i -g @openai/codex          && codex login
npm i -g @google/gemini-cli     && gemini
# claude CLI: https://docs.claude.com/en/docs/claude-code/quickstart

# 2. Install as a Claude Code skill
cp -r harness ~/.claude/skills/mixture-of-agents

# 3. Inside Claude Code, in any project
/mixture-of-agents

Architecture at a glance

Layer 0 — Scout brief           (parent Claude, in-place)
Layer 1 — Proposers (3 parallel)  codex + gemini + sonnet subprocesses
Layer 2 — Broadcast refiners (2)  codex + gemini, each sees ALL 3 proposals
Layer 3 — Aggregator              (parent Claude Opus, in-place)

Typical wall-clock is 6–12 minutes. Use it for non-trivial architecture work, not one-line fixes. Background in docs/architecture.md.

Docs

docs/install.md: install the three CLIs, verify, install as a Claude Code skill
docs/usage.md: running via /mixture-of-agents (primary) or standalone
docs/config.md: .env + harness/config.yaml, MOA_* knob table, precedence
docs/architecture.md: the four layers, why broadcast, why these three providers
CONTRIBUTING.md: dev setup, PR protocol, where help is welcome
SECURITY.md: private vulnerability reports
CLAUDE.md: guidance for coding agents working on this repo

Repo layout

README.md              this file
CLAUDE.md              agent guidance for this repo
CONTRIBUTING.md        contributor guide
SECURITY.md            vulnerability reporting
LICENSE                MIT
.env.example           copy to .env to override harness defaults
docs/                  longer-form docs by topic
harness/               orchestrator, adapters, prompts, schemas
  SKILL.md             Claude Code skill manifest
  README.md            skill-internal notes (lives with harness/ when copied into ~/.claude/skills/)
  config.example.yaml  copy to harness/config.yaml to override defaults
  prompts/             scout / proposer / refiner / aggregator
  scripts/             orchestrator + adapters + config + tests
requirements-cli.txt   install/auth notes for the three CLIs

PRs we'd love to see

The current harness is shaped around what I use day-to-day. These are the directions that would expand who MoA-X is useful for, and I'd prioritize reviewing PRs that land any of them:

API-billing support for codex, gemini, and claude. Right now the adapters assume each CLI is logged in against a subscription plan. Shops that run through ANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY should be a first-class path.
OpenCode and other agent harnesses. The orchestrator runs fine from a plain shell, but the scout + aggregation steps are tailored to Claude Code. PRs that close the gap for OpenCode (or aider, codex-as-harness, roo, continue, cline, etc.) are welcome.
Chinese-lab models. DeepSeek, Qwen, Kimi, GLM, MiniMax. The whole argument for MoA is cross-lab diversity, and the current lineup is US-only. Routing one proposer through a Chinese frontier model would test the thesis in the place it most deserves testing.
More providers generally. xAI Grok, Mistral, any model with a credible coding-bench story. Each provider needs its own adapter, preflight, and prompt-assumption review; open an issue first so we can talk through auth shape, then build.
Cost observability for API-billed runs: token accounting in the manifest, a MOA_MAX_COST ceiling, per-layer spend breakdowns.

See CONTRIBUTING.md for the PR protocol.

Status

Early open-source release. Contributions welcome; see CONTRIBUTING.md. Security reports go through SECURITY.md.

License

Author

Kyle Boddy.