OpenAI Symphony is an agent orchestrator that uses project-management tools, like issue trackers, as a control plan to coordinate multiple coding agents. Instead of developers managing interactive coding sessions, Symphony manages "tasks" by assigning each one to a dedicated agent that works autonomously to completion. Once a task is finished, a human is in charge to review the resulting output.
OpenAI engineers created Symphony to overcome the "human attention" bottleneck they were experiencing using a more naïve workflow:
Each engineer would open a few Codex sessions, assign tasks, review the output, steer the agent, and repeat. In practice, most people could comfortably manage three to five sessions at a time before context switching became painful.
Beyond that point, it became harder for engineers to remember which session was doing what, monitor stalled agents, and maintain a clear mental model of ongoing work.
So, instead of organizing work around individual coding sessions, each aiming at eventually merging a PR under explicit human supervision, Symphony structures its workflow around the core deliverables of a project as its workflow building blocks, including issues, tasks, tickets, and milestones.
Symphony continuously watches the task board and ensures that every active task has an agent running in the loop until it’s done. If an agent crashes or stalls, Symphony restarts it. If new work appears, Symphony picks it up and starts organizing work.
In this model, agent work is not longer tied to PRs. An issue might instruct an agent to analyze the codebase and generate an implementation plan, then break it down into a tree of tasks that Symphony can schedule across other agents. Similarly, if an agent identifies an opportunity for optimization or refactoring, it can open a new issue on its own. In both cases, a human developer remains responsible for reviewing these generated issues before Symphony assigns them for execution.
The main advantage of this approach is that the cost of an agent making mistakes is significantly reduced, since it primarily involves reviewing completed work and rejecting it.
Another notable aspect of Symphony is that it is not a complex supervision system, but rather a SPEC.md file describing the problem and its solution that every organization can use to create its own orchestrator. The reference implementation is built in Elixir for its "excellent primitives for orchestrating and supervising concurrent processes".
As a final note, it is worth noting that OpenAI does not position Symphony as a standalone product. Instead, it is a reference implementation that developers can adapt and tailor to their own use cases and repositories.