muse-agents

Harness-independent AI agent definitions for running a web agency — 21 core agents across 8 divisions, plus 13 opt-in packs (49 total agents).

What is this?

muse-agents is a portable roster of agency AI agents: business development, marketing, SEO, design, development, operations, quality gate, and evidence fact-checking. Each agent is a thin persona — trigger words, pre-flight checks, workflow, output contract, handoffs — while domain knowledge lives separately in muse-skills. Definitions are harness-neutral Markdown rendered into your tool of choice (OpenCode, Codex TOML, and Claude Code adapters included).

Key Architectures & Packs

• Gauntlet Multi-Agent Loop (packs/gauntlet/): Bounded iterative improvement engine (Judge freezes contract -> Builder writes minimal diff -> Critic fresh review with review.json -> Judge evaluates plateau -> Integrator merges single largest fix).
• Governed Secretary & Staff Work (agents/ops/secretary.md & packs/secretary/): Evidence-grounded staff-work synthesizer enforcing mandatory Dissent, Uncertainties & Omissions sections, primary citation matrices, and SHA-256 action approval tokens.
• Source & Claim Fact-Checker (agents/research/fact-checker.md): Adversarial citation and evidence auditor tagging claims with [RAW], [FETCH], [SEARCH], and [INFER].
• Torvalds Code Reviewer (agents/dev/code-reviewer.md): Enforces binary correctness, data-structure elegance, zero regressions, and minimal complexity patches.

Quick Start

# clone
git clone git@github.com:harshsinghmp/muse-agents.git
cd muse-agents

# validate the roster (no dependencies beyond bash + python3 / bun)
scripts/build.sh --check

# install into OpenCode (~/.config/opencode/agents/)
scripts/install.sh --core                         # opencode (default), 21 core agents
scripts/install.sh --harness codex --all          # codex TOML agents (49 total)
scripts/install.sh --packs gauntlet,secretary     # selected packs

Dispatch by trigger words listed in each agent's Role section. Code work defaults to backend-arch triage. Every deliverable passes nexus-reviewer before client-facing release.

Other harnesses (Claude Code renderer, Codex TOML, Hermes ingestion): adapters/README.md.

Project Structure

muse-agents/
├── adapters/
│   ├── claude-code/
│   ├── codex/
│   └── opencode/
├── agents/
│   ├── bizdev/
│   ├── design/
│   ├── dev/
│   ├── gate/
│   ├── marketing/
│   ├── ops/
│   ├── research/
│   └── seo/
├── packs/
│   ├── audit/
│   ├── ecommerce/
│   ├── gauntlet/
│   ├── hosting-ops/
│   ├── legal/
│   ├── mobile/
│   ├── outreach/
│   ├── pr/
│   ├── product/
│   ├── research/
│   ├── secretary/
│   ├── video/
│   └── wordpress/
├── scripts/
├── spec/
├── tests/
├── AGENTS.md
├── README.md
├── STATE.md
└── SUMMARY.md

Documentation

Resource	Description
AGENTS.md	Framework contract: dispatch rules, subagent discipline, quality gate
spec/agent-format.md	Canonical schema every definition must satisfy
adapters/opencode/README.md	Renderer usage for OpenCode subagents
docs/research/aidevops-mining.md	Pattern research behind the format

Contributing

New agents must satisfy spec/agent-format.md and pass scripts/build.sh --check before merge. Add core agents under agents/<division>/, optional specialists under packs/<name>/.

License

MIT — see LICENSE.

---