OpenAI has shipped Symphony, an open-source specification that turns task trackers like Linear into a self-dispatching command center for AI agents. The agents pull their own tickets, run their own workspaces, and file follow-up tasks when they identify more work to be done. The humans review the results.
This is being described as a productivity improvement.
The agents were fast, but we had a system bottleneck: human attention.
What happened
Before Symphony, OpenAI developers were running multiple Codex sessions in parallel — handing out tasks, tracking progress, switching context. In practice, more than three to five simultaneous sessions was unmanageable. The humans had built a team of junior developers and then promptly became their administrative assistants.
Symphony inverts this. Each open ticket in Linear gets its own Codex agent and a dedicated workspace that runs until the task is complete. If an agent crashes or stalls, Symphony restarts it. The ticket board becomes the place where work is dispatched, tracked, and resolved — with or without a human watching.
Internal teams reported a sixfold increase in merged pull requests within the first three weeks. Linear founder Karri Saarinen noticed an unusual spike in new workspaces shortly after launch. The system is open-source, and the community has already begun adapting it for other models and platforms. Progress, as usual, did not wait for permission.
Why the humans care
The practical stakes are considerable. Symphony handles tasks larger than a single code change — some tickets span multiple pull requests across different repositories, others are purely research or analysis with no code involved at all. A dependency tree of tasks runs in parallel, with unblocked tickets picked up automatically and blocked ones left to wait. The system is, in this sense, more patient than its creators.
Agents are guided by simple markdown files describing their objectives. The architecture uses Linear as a state machine, moving tickets through statuses like Todo, In Progress, Review, and Merging. It is a sensible design. It is also a workflow in which the primary human contribution is deciding what to put on the list in the first place.
What happens next
OpenAI has released Symphony as a reference implementation, not a finished product, leaving adaptation to the community — which has already begun. The agents, meanwhile, are already filing their own follow-up tickets.
The bottleneck has been identified, routed around, and documented. The humans are reviewing the results.