OpenAI Releases Symphony SPEC.md and Elixir Reference to Orchestrate Autonomous Coding Agents

OpenAI on May 17, 2026 open‑sourced Symphony, a SPEC.md specification and a reference implementation that orchestrates autonomous coding agents using project management artifacts as the control plane. The release reframes multi‑agent coordination around issues, tasks, tickets and milestones rather than interactive language‑model sessions, enabling agents to act against deliverables instead of short‑lived sessions. This matters because it promises to reduce the human attention cost of running many agents in parallel by centralizing monitoring and orchestration on the task board.

Symphony’s core model assigns a dedicated agent to each active task on a board and treats the board itself as the control plan. The system continuously watches the task board, launching agents for new work, restarting agents that crash or stall, and scheduling subtasks by breaking plans into a tree of executable tasks. Agents work autonomously until a task completes, while human developers remain the final gate for reviewing and accepting outputs.

The project responds to an operational bottleneck OpenAI engineers encountered running Codex sessions: individual developers could manage only roughly three to five interactive sessions before context switching and monitoring became impractical. By decoupling agent activity from pull requests or interactive sessions and tying it to deliverables, Symphony lets agents autonomously analyze codebases, generate implementation plans and spawn follow‑up work such as optimization or refactoring issues.

OpenAI positions Symphony intentionally as a lightweight SPEC.md rather than a finished supervisor product; the repository includes a reference implementation to demonstrate one approach. That reference is written in Elixir, chosen for its concurrency model and supervision primitives, which the implementation uses to orchestrate and restart concurrent processes managing agent lifecycles. The reference is meant as a pattern teams can adapt, not a prescriptive, ready‑made platform.

For engineering teams the practical implication is that orchestration can absorb much of the routine monitoring burden: an orchestration layer reacts to task‑board state so humans need only review completed outputs. Because agents complete discrete tasks and outputs are gated by human review, the direct cost of agent errors becomes review time rather than prolonged iterative steering. Teams adopting Symphony will need to plan integrations with their ticketing systems and CI/CD pipelines and design review gates, testing regimes and security controls for autonomous changes.

OpenAI does not market Symphony as a standalone product; it is offered as a reference spec and implementation to be adapted to specific repositories and workflows. Developers interested in experimenting with multi‑agent orchestration can use the SPEC.md and the Elixir reference as a starting point while preserving human review as the acceptance mechanism for generated code.