OpenAI Open-Sources Symphony: Your Linear Board Is Now an Agent Factory

OpenAI released Symphony, an open-source specification and Elixir reference implementation that turns project management boards into control planes for coding agents. Connect it to Linear, and every open task gets an agent. Agents run continuously until PRs merge. Crashed agents auto-restart. OpenAI reports a 500% increase in landed PRs in the first three weeks of internal use.

That number is wild, and I'm not sure I fully trust it without more context about what "landed PRs" means at OpenAI's scale. But the architecture is genuinely interesting. Symphony treats your issue tracker as the source of truth for agent work. You don't write agent configurations. You write tickets. The orchestrator reads them and dispatches.

Version 1.1.0 adds Kata CLI support, which makes Symphony model-agnostic. Claude Code, Gemini, and other models can run inside the same orchestration framework. This is important because it separates the orchestration layer from the model layer. You pick the best model for each task type, and Symphony handles the lifecycle.

The pattern here, "project board as agent control plane," is a genuinely new primitive. I've been thinking about this for weeks as I watch my own MindPattern pipeline evolve. The bottleneck in agentic coding isn't the model. It's the orchestration. Who decides what work to do next? Who monitors whether the agent is stuck? Who restarts it when it crashes? Symphony answers all three with: your existing project management workflow.

This pairs with Guillermo Rauch's Open Agents announcement, where the Vercel CEO open-sourced a full reference platform for building cloud coding agents. Rauch's framing: "the moat of software companies will shift from the code they wrote to the means of production of that code." Companies like Stripe (Minions), Ramp (Inspect), Spotify (Honk), and Block (Goose) have been building exactly this internally. Now the reference implementations are public.

For builders, the action item is concrete. If you're running any kind of agent-assisted development, look at Symphony's architecture even if you don't use Elixir. The "every ticket gets an agent" pattern will become standard within a year. Start structuring your tickets so an agent can parse them: clear acceptance criteria, specific file paths, explicit test requirements. The agents are only as good as the work items you feed them.