OpenAI just open-sourced Symphony, a specification that turns tools like Linear into orchestration layers for coding agents. Every open task gets its own agent that runs until the job is done. No more babysitting sessions. Some OpenAI teams saw a 500% jump in landed pull requests within three weeks, according to the project's authors Alex Kotliarskyi, Victor Zhu, and Zach Brock.
The system came out of an internal experiment where OpenAI built an entire repository with zero human-written code. Every line came from Codex. It worked, but engineers hit a wall fast. Managing multiple agent sessions meant constant context switching. People could handle maybe three to five sessions before things fell apart. Symphony fixes this by making the issue tracker the control plane. Each ticket becomes a state machine. Agents pull work, break it into stages, file new issues when they spot improvements. One engineer shipped three significant changes from a cabin using just the Linear phone app.
But there are real tradeoffs. When you stop steering agents interactively, you lose the ability to course-correct mid-flight. Sometimes agents miss the mark entirely. The Symphony team's response was to build better guardrails rather than patch failures by hand. It's a bet that upfront investment in constraints pays off more than constant human oversight.
For the agent space, this is one of the first concrete specs for coordinating multiple coding agents at scale. Linear founder Karri Saarinen noted a spike in workspaces created when Symphony launched. If the pattern holds, project management tools become the natural interface for agentic development, not terminals or chat windows. That's a big shift in how we think about who, or what, actually writes code.