OpenAI Publishes Symphony Spec

To solve the context switching problem that arose in internal Codex workflows, OpenAI has released an open-source spec for Symphony, an agent orchestrator that uses an issue tracker as its control plane.

For every open ticket in Linear, it attaches a dedicated workspace and agent, and relaunches them if a crash or stall occurs. Work is organized by ticket rather than by session, and a single issue may be split across multiple PRs or end with just an analysis task.

• Dependencies are managed like a DAG, running unblocked tasks in parallel.
• Agents can also create new issues or propose follow-up improvements.
• At some teams, the number of merged PRs increased by 500% during the first 3 weeks.

PMs and designers can also submit work without touching the repository directly, and the system handles CI monitoring, rebasing, conflict resolution, and retrying flaky checks. In exchange, fine-grained mid-process steering has decreased, which required strengthening guardrails, tests, Chrome DevTools, and QA smoke tests, and the approach is now shifting toward giving goals and tools rather than strict state transitions. The heart of the repository is not complex code but specs like SPEC.md and WORKFLOW.md.