Symphony

Symphony is an open-source orchestration system from OpenAI that converts project-board tasks into isolated, autonomous coding runs. It monitors a task board such as Linear, spawns Codex agents to handle each task, collects proof-of-work evidence (CI status, PR review feedback, complexity analysis, and walkthrough videos), and lands approved pull requests automatically. The goal is to let engineers manage work rather than supervise individual coding agents.

There are two paths. First, you can use the Elixir-based reference implementation included in the repository under `elixir/README.md`, which walks through environment setup. Second, you can point any coding agent at SPEC.md in the repo and ask it to build Symphony in the language of your choice. The README also suggests asking a coding agent to help with setup using the provided instructions.

Yes, Symphony is licensed under the Apache License 2.0 and published publicly by OpenAI on GitHub. The source code, spec, and Elixir reference implementation are all freely available. Note that running the underlying Codex agents that Symphony orchestrates may incur API costs separately.

Symphony is best for engineering teams that have already invested in harness engineering, meaning their codebase has reliable CI pipelines, tests, and lint checks. In that context, Symphony can spawn agents against task-board items and use those harnesses as meaningful proof-of-work signals. Teams managing high volumes of well-defined, bounded implementation tasks are the primary fit.

Running a coding agent directly requires an engineer to supervise each step: prompting, reviewing intermediate outputs, and deciding when to commit. Symphony adds an orchestration layer that automates the full cycle from task intake to PR merge, with structured proof-of-work artifacts replacing ad-hoc supervision. Per the README, the shift is from 'managing coding agents' to 'managing work that needs to get done,' which is a meaningful abstraction jump.

The README explicitly warns that Symphony is a 'low-key engineering preview for testing in trusted environments,' signaling it is not production-hardened. Its proof-of-work model depends entirely on the quality of a codebase's existing harnesses; weak or absent CI and tests mean the agents' output cannot be meaningfully verified before merge. It should not be run in untrusted or shared infrastructure at this stage.