Open-Source Spec for Orchestration: Symphony

OpenAI built Symphony, which uses an issue tracker as its control plane, to move beyond the limits of directly conducting Codex by hand. The core idea is to treat each Linear issue as an agent's unit of work, attaching an agent to every open task and keeping it running continuously.

The initial problem was context switching rather than productivity. Once an engineer went beyond 3-5 sessions at a time, it became confusing who was doing what, and long-running tasks were prone to stalling or drifting off track partway through.

Symphony turned this into a ticket-based state machine. Agents work by following issue state, restart when interrupted, and immediately pick up new work as it appears. It's a structure that breaks work down around issues, tasks, and milestones rather than being PR-centric.

With this approach, some teams saw landed PRs increase by 500%. A more important change is that the cost of exploration dropped sharply: engineers can easily throw out experimental refactors or hypothesis tests, and PMs and designers can also submit feature requests directly and get back review packets that work in the actual product.

Operationally, Symphony is also designed for large monorepos. It handles CI monitoring, rebasing, conflict resolution, and retrying flaky checks, increasing the likelihood that once a ticket reaches the Merging stage, it lands on the main branch without repeated human intervention.

The implementation itself starts from a surprisingly simple SPEC.md. Policies and prompts are managed through a repo-owned WORKFLOW.md, and stricter approval and sandbox policies can be chosen per implementation as needed. Ultimately, the direction Symphony points toward is treating agents less like fixed state-transition machines and more like teammates given goals.