An open-source spec for orchestration: Symphony

Every engineering team has the same quiet problem: the backlog never stops growing, and the humans assigned to it keep getting interrupted. Symphony is OpenAI's answer to that, a specification released as open source that wires Codex agents directly into issue trackers so tickets can be picked up, worked, and pushed toward resolution without someone babysitting the process.

The idea is less about a single clever bot and more about plumbing. Symphony defines how agents should read tasks, claim them, report progress, and hand off work, all through the same interface engineers already use to track bugs and features. Instead of a developer copy-pasting a ticket into a chat window and waiting around, the orchestration layer treats the tracker itself as the job queue. Agents pull from it continuously, which is where the 'always-on' framing comes from.

OpenAI's own numbers, drawn from internal use, point to fewer interruptions and more throughput on engineering work, though the company is careful to frame this as a specification rather than a finished product. That distinction matters: Symphony isn't a single hosted service you subscribe to, it's a set of conventions other teams and tools can implement, extend, or bolt onto existing systems like Jira or Linear. Open-sourcing the spec rather than just shipping a feature suggests OpenAI wants this to become a shared standard for agent orchestration, the same way something like OpenAPI became a lingua franca for describing REST services.

What's notable is the framing around context switching specifically. A lot of AI coding tools sell speed on a single task. Symphony is pitched at the layer above that, the messy reality of juggling dozens of open tickets, priorities shifting mid-sprint, and engineers losing time just remembering where they left off. If agents can hold that state instead of humans, the argument goes, the humans get to spend more of their attention on actual engineering decisions rather than bookkeeping.

Whether other companies adopt the spec, fork it, or ignore it in favor of their own orchestration layers is the real test here. Specs only matter if people build against them.