OpenAI just published a blog post detailing the internal impact of Symphony, its agent orchestration tool. Symphony, which was open-sourced earlier this month, uses Linear’s task board as a dispatch center for programming agents: each to-do item gets its own dedicated agent and isolated workspace, and the agent keeps running until the task is done or triggers a human review. After some internal teams adopted it, their merged PR count jumped 500% in three weeks.
The post also reveals some previously undisclosed usage details. Product managers and designers can now submit feature requests directly in Symphony without touching the codebase or managing a Codex session. The review packet they get back includes a video of the feature running in the actual product. Agents can also break complex tasks into multi-stage dependency trees — for example, finishing a Vite migration before starting a React upgrade. Any improvement opportunities the agent spots during implementation automatically spawn new tasks, which humans evaluate before scheduling.
Symphony is, at its core, a single spec file — SPEC.md. To verify the spec’s completeness, the team had Codex implement it in TypeScript, Go, Rust, Java, and Python. Each one worked. The reference implementation ended up in Elixir because its concurrency primitives fit the orchestration use case. As the team put it, "When code is essentially free, you can finally choose a language based on its strengths rather than ecosystem inertia."
OpenAI says it doesn’t plan to maintain Symphony as a standalone product. Instead, it’s a reference implementation — the company encourages developers to use its spec to build their own version with any programming agent.