Symphony

Orchestrate your code.

Symphony is an AI agent framework that dispatches a coordinated team of specialist agents to deliver software products. Describe your goal — Symphony picks the phase, the workflow, and the agents.

It runs inside your AI coding tool of choice. No standalone CLI. No new binary. Just your existing workflow, supercharged.

Supported Platforms

| Platform | Config File | Commands | |----------|------------|----------| | Claude Code | CLAUDE.md | .claude/commands/symphony-*.md | | GitHub Copilot | .github/copilot-instructions.md | .github/prompts/symphony-*.prompt.md | | OpenAI Codex | AGENTS.md | .codex/symphony-*.md |

What Makes Symphony Different

Smart orchestration — You describe what you want; the Conductor picks the lifecycle phase, workflow, and agents. No command memorization.

Parallel execution — Agents run concurrently in waves with integration gates between waves, producing multi-step outputs faster than any sequential framework.

Dual-layer architecture — A tool-agnostic Core runtime with thin adapters for each platform. Future AI tools plug in through new adapters without touching Core.

Engine-enforced disciplines — Self-critique thresholds, knowledge trust levels, anti-rationalization tables, and diagnose-then-fix recovery are baked into the engine, not opted into per agent.

Framework at a Glance

| | Count | |---|---| | Agents | 35 | | Workflows | 76 | | Protocols | 21 | | Knowledge Fragments | 51 | | Skills | 12 | | Adapters | 3 (Claude Code, Copilot, Codex) | | Tests | 1700+ |

The 5-Phase Lifecycle

Analysis → Planning → Solutioning → Implementation → Deployment

Each phase has dedicated workflows and agents. The Conductor routes you to the right phase based on your goal. Anytime workflows (quick-spec, quick-dev, spike, explore-codebase, etc.) are available in any phase.

Architecture

_symphony/
├── _config/          Global config, manifest, presets
├── core/             Engine specs, protocols, vault, CLI registry
├── lifecycle/        5-phase workflows + agents
├── dev/              Developer agents, skills, knowledge
├── creative/         Creative agents + workflows
├── testing/          Test agents, workflows, knowledge
├── utility/          Cross-cutting utility agents
└── _memory/          Checkpoints + sidecars

adapters/
├── claude-code/      Claude Code adapter
├── copilot/          GitHub Copilot adapter
└── codex/            OpenAI Codex adapter

See docs/superpowers/specs/2026-04-08-symphony-architecture-design.md for the full architecture spec.

Getting Started

• Getting Started with Claude Code
• Getting Started with GitHub Copilot
• Getting Started with OpenAI Codex

Status

v0.1.0 — Specification complete. All agents, workflows, protocols, and knowledge fragments are fully specified. The runtime engine that executes these specifications is planned for v0.2.0.

What's included in v0.1.0:

• Complete agent persona definitions with expertise, knowledge trust boundaries, and anti-rationalization pairs
• Full workflow specifications (YAML metadata + XML instructions + Markdown templates)
• 21 protocols covering cross-cutting concerns (self-critique, checkpoints, memory hygiene, workspace isolation, etc.)
• 3 working adapter translators that generate platform-native command files
• 1700+ structural validation tests
• Vault knowledge graph schema for codebase intelligence
• Environment presets for solo, team, and enterprise configurations

Running the Tests

npm install
npm test

Contributing

See CONTRIBUTING.md for how to add agents, workflows, adapters, and knowledge fragments.

License

AGPL-3.0. See LICENSE.