OpenAI open-sources Symphony spec

OpenAI published Symphony on April 27 as an open-source specification for coordinating autonomous coding agents through project-management tools rather than one chat or terminal session at a time. The company described the system as a way to turn an issue tracker such as Linear into a control plane that assigns work to dedicated agents, keeps those agents running, and routes the results back for human review. A GitHub repository for the project includes a `SPEC.md` file and an experimental reference implementation written largely in Elixir. OpenAI said the approach produced a 500% increase in landed pull requests on some internal teams. ### Why did OpenAI build a spec instead of another coding assistant? OpenAI said Symphony grew out of an internal workflow problem that appeared after teams started using Codex heavily. In a post by Alex Kotliarskyi, Victor Zhu and Zach Brock, the company said engineers could usually manage only three to five concurrent coding sessions before context switching became a bottleneck. The OpenAI post said the company changed the unit of work from an agent session or pull request to a task tracked in software teams’ existing systems — issues, tickets and milestones. (openai.com) Under that design, the human role shifts from steering each agent continuously to reviewing completed work and approvals at the task level. ### What does Symphony actually specify? The GitHub repository says Symphony “turns project work into isolated, autonomous implementation runs,” with each task mapped to its own workspace and agent process. (openai.com) In OpenAI’s description, the issue tracker becomes the system of record: every open task gets an agent, agents run continuously, and humans review the outputs. InfoQ reported that the spec centers on `SPEC.md` rather than a single packaged product. The repository README says developers can either implement the specification in another language or use OpenAI’s experimental Elixir version, which the company positioned as a reference implementation rather than a standalone commercial offering. (openai.com) ### How does the workflow differ from a normal agent session? OpenAI said Symphony monitors a Linear board for work, spawns agents to handle tasks, and keeps those runs isolated from one another. (github.com) The repository README says agents can return “proof of work” including continuous-integration status, pull-request review feedback, complexity analysis and walkthrough videos before a change is accepted and landed. InfoQ said the model also allows an issue to produce more issues. (infoq.com) In the publication’s description of the spec, an agent can analyze a codebase, generate an implementation plan, break it into a tree of tasks and schedule that work across other agents, while human reviewers remain responsible for approving generated issues before execution. ### What does this change about evaluating agent systems? InfoQ said Symphony pushes orchestration toward task decomposition, handoffs and reviewable workflow state rather than a single interactive exchange. (github.com) That means the records left behind by the system — issue updates, approvals, CI results and other artifacts — become part of how teams judge whether the process worked. OpenAI’s own materials stop short of laying out a formal benchmark for those workflows, but the structure described in the spec points toward evaluation at the project level: whether tasks were split correctly, whether agents handed work off cleanly and whether reviewers had enough evidence to approve or reject results. (infoq.com) That is an inference from the published design and repository materials, not a separate OpenAI statement. ### Who can use it now, and what are the limits? The GitHub repository is public and licensed under Apache 2.0, and the README says teams can adapt the specification to their own repositories or ask another coding agent to implement it in a language of their choice. The same README calls the current release a “low-key engineering preview for testing in trusted environments,” which signals that OpenAI is not presenting it as a finished general-purpose product. (openai.com) GitHub showed the `openai/symphony` repository live as of May 18, 2026, with the spec file, an Elixir implementation directory and no published releases. OpenAI’s post links directly to that repository, where the next steps for users are to read `SPEC.md` or follow the Elixir setup instructions in the project documentation. (github.com)