OpenAI's Symphony suggests spec.md
- OpenAI published Symphony on April 27, an open-source orchestration spec that turns task boards like Linear into control planes for Codex coding agents. - OpenAI said Symphony delivered a 500% increase in landed pull requests on some teams, after engineers hit limits managing three to five agents. - The release extends OpenAI’s “no human-written code” workflow into task-level automation and human review loops. (openai.com)
OpenAI said this week that Symphony is the next layer in its agent-first coding stack: a written spec that assigns coding agents to open tasks. (openai.com) The company published Symphony on April 27 and described it as an open-source orchestration spec for Codex. In OpenAI’s setup, a project-management board such as Linear becomes the control plane, and each open issue maps to a dedicated agent workspace. (openai.com) OpenAI said the system came after engineers hit a limit with interactive coding sessions. The company wrote that most people could manage only three to five Codex sessions at a time before context switching started to drag productivity down. (openai.com) The basic idea is simple: write down the job once, then let software keep pulling from that description until a human reviews the result. OpenAI said every open task gets an agent, agents run continuously, and humans review the output instead of micromanaging each session. (openai.com) That pushes the “spec” closer to the center of the workflow than the code editor. OpenAI wrote that it had been optimizing around sessions and merged pull requests, then shifted to deliverables such as issues, tasks, tickets, and milestones. (openai.com) Symphony also builds on an earlier OpenAI experiment published on February 11. In that post, OpenAI said a small team built an internal beta with zero lines of manually written code, about a million lines generated by Codex, and roughly 1,500 pull requests merged over five months. (openai.com) That earlier project framed the engineer less as a person typing code and more as a person designing tests, repository rules, and feedback loops. OpenAI estimated the product was built in about one-tenth the time it would have taken to write the code by hand. (openai.com) Symphony is OpenAI’s answer to the next bottleneck in that model: human attention. The company said some teams saw a 500% increase in landed pull requests after moving from manually supervised agent sessions to task-driven orchestration. (openai.com) The release also sharpens a pattern developers have been circling for months: put the durable instructions in files such as AGENTS.md or a project spec, and let agents handle implementation details inside guardrails. OpenAI’s February post said even the initial AGENTS.md in its repo was written by Codex itself. (openai.com) For OpenAI, the closing argument is operational, not philosophical. If the task tracker becomes the control plane, the human reviewer becomes the scarce resource Symphony is designed to protect. (openai.com)
