OpenAI’s Symphony is a new open-source spec and reference implementation for orchestrating coding agents around a project-management board such as Linear. The company describes it as a control plane in which every open task gets an agent, the agents run continuously, and humans review the results.
The post, published on April 27, argues that coding agents hit a human bottleneck once teams start supervising multiple sessions at once. OpenAI portrays Symphony as a response to that problem: rather than managing Codex sessions directly, the workflow shifts to issues, tickets, milestones, and other deliverables already used to organize software work.
According to the company, Symphony maps each open Linear issue to a dedicated workspace and keeps agents running until tasks are complete or no longer eligible. OpenAI also claims that, on some teams, the approach produced a "500% increase in landed pull requests" in the first three weeks. As with any internal productivity metric, that figure should be treated cautiously.
The system is designed to do more than feed issues into agents. OpenAI says agents can create follow-up work, split large tasks into dependency trees, and continue through multiple turns on the same thread when an issue remains active. Symphony also watches CI, rebases changes when needed, and handles flaky checks and merge-related cleanup in large monorepos.
A substantial part of the post is devoted to the spec itself. OpenAI presents Symphony as a `SPEC.md`-driven system that defines the workflow contract, configuration, orchestration state machine, workspace rules, and integration boundaries. The company says the reference implementation is written in Elixir, and that it asked Codex to implement the spec in several other languages — including TypeScript, Go, Rust, Java, and Python — to expose ambiguities and simplify the design.
OpenAI also highlights Codex App Server as a core building block. The company says Symphony uses Codex in headless app-server mode and relies on JSON-RPC-style messaging for starting threads and reacting to turns. To avoid exposing Linear access tokens to subagents, OpenAI mentions using dynamic tool calls to expose a raw `linear_graphql` function instead of relying on MCP.
The post presents Symphony as intentionally minimal and not as a standalone product the company plans to maintain long term. OpenAI encourages others to point their own coding agents at the spec and repository to build versions tailored to their own environments. It also notes that the repository had gathered over 15K GitHub stars as of April 23, a figure that appears to reflect early developer interest more than a settled verdict on the system itself.
Source: OpenAI
