Symphony¶

What it is¶

Symphony is an autonomous implementation framework by OpenAI (July 2026) that turns project work into isolated, autonomous implementation runs. It manages high-level work items (issues) by coordinating a fleet of coding agents that handle the end-to-end implementation lifecycle using the MCP 3.0 Task Protocol.

What problem it solves¶

It solves the "supervision bottleneck" in agentic software engineering by shifting the human role from direct code-level supervision to high-level work item management. By utilizing the MCP 3.0 Task Protocol for standardized multi-agent coordination, Symphony enables a "software factory" model where agents operate with high autonomy and verifiable output.

Where it fits in the stack¶

Agents / Orchestration Framework. It manages the lifecycle of implementation runs, interfacing between task trackers (e.g., Linear, Jira) and version control systems (GitHub, GitLab), while using Model Context Protocol (MCP) for tool and task standardization.

Typical use cases¶

Multi-Agent Coordination: Using the MCP 3.0 Task Protocol to distribute complex implementation tasks across specialized agents.
Issue-to-PR Automation: Automating the transition from a descriptive issue to a verified Pull Request.
Fleet Management: Managing dozens of agents working in parallel on isolated features, bug fixes, or documentation updates.
Continuous Implementation: Integrating autonomous coding agents into the CI/CD pipeline to handle routine maintenance tasks.

Strengths¶

MCP 3.0 Task Protocol: Leverages standardized task representations for seamless coordination between diverse agent types.
Autonomous Implementation Runs: Spawns agents to handle tasks from a work board without constant human prompting.
Proof of Work: Enforces standardized verification (CI status, walkthrough videos, PR review feedback).
Safe Landing: Built-in mechanisms for safely merging PRs once automated and manual checks pass.

Limitations¶

Harness Dependency: Requires robust testing environments and CI to ensure agentic changes don't introduce regressions.
Experimental State: Many implementations are still in the reference stage (e.g., the Elixir-based reference implementation).
Context Management: Can be token-intensive when processing large repositories or complex specifications.

When to use it¶

When you want to move from micro-managing coding agents (like Aider) to managing high-level work items.
In mature codebases that have already adopted Software Factories principles.
When you need to scale engineering capacity using GPT-5.5 or Claude 4.8 agents.

When not to use it¶

In small projects where manual agent supervision is not a bottleneck.
In environments where comprehensive testing and CI are not yet established.
When working on legacy systems that lack automated test suites.

Getting started¶

Requirements¶

A codebase with established CI and testing harnesses.
API access to a frontier model (Claude 4.8 Opus or GPT-5.5).
An MCP 3.0 compatible environment for task coordination.

Installation¶

Clone the official repository to explore the reference implementation:

git clone https://github.com/openai/symphony.git
cd symphony

For the Elixir reference implementation:

cd elixir
mix deps.get
mix compile

Basic Implementation Run¶

# Configure the symphony environment
export SYMPHONY_MODEL=gpt-5.5-preview
export SYMPHONY_MCP_ENDPOINT=http://localhost:8000/task-protocol

# Start an implementation run for a specific issue
symphony run --issue BUG-404 --verify-with-ci

CLI examples¶

# Start the symphony service with a specific workflow file
symphony start --workflow ./WORKFLOW.md

# Check the status of active implementation runs across the fleet
symphony status --detailed

# Manually trigger the MCP 3.0 Task Protocol handshake
symphony mcp handshake

API examples¶

Symphony provides an HTTP API for monitoring and management (typically at /api/v1/state):

{
  "state": "active",
  "mcp_version": "3.0",
  "active_runs": [
    {
      "issue_id": "FEAT-202",
      "agent_type": "claude-4-8-opus",
      "task_protocol_status": "in_progress",
      "ci_status": "passing"
    }
  ]
}

Sources / references¶

Contribution Metadata¶

Last reviewed: 2026-07-21
Confidence: high