OpenAI Releases Symphony for Autonomous AI Agents
OpenAI introduced Symphony, an open-source specification complete with a reference implementation. This tool converts task trackers such as Linear into hubs for Codex agents. Developers no longer need to handle multiple sessions at once. Agents select open tickets on their own, and people only check the outcomes.
Internal groups at OpenAI reported a sixfold rise in merged pull requests during the first three weeks of use. Karri Saarinen, founder of Linear, observed more new workspaces created in the tool following the launch.
Each open ticket receives its own Codex agent and workspace. The agent operates until completion, making the task board the spot where assignments happen.
Human Attention Limits Prompt Change
Prior to Symphony, OpenAI staff managed various Codex sessions concurrently. They assigned tasks and monitored advancement in each. Handling over three to five sessions proved difficult, as frequent switches reduced efficiency.
Developers noted, "The agents were fast, but we had a system bottleneck: human attention." They had assembled a group of junior developers, yet burdened colleagues with oversight duties. This led to reversing the process. Agents now draw work directly from the tracker, eliminating supervision of individual sessions.
Linear Serves as Core State Machine
Symphony treats Linear as a state machine. Tickets progress through stages including "Todo," "In Progress," "Review," and "Merging." The setup monitors the board to ensure every active ticket has an agent. If an agent fails or pauses, the system restarts it.
Only tickets without blocks get assigned, supporting parallel task branches. For example, a React upgrade began after a Vite migration upstream. Tickets often exceed single code edits. Some generate multiple pull requests over repositories, while others involve research or analysis without code.
Agents detect unrelated issues, such as performance flaws or refactoring needs, and create new tickets. Product managers and designers submit feature requests directly. They receive review packages with video demos, bypassing repository access.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Symphony handles tasks of varying complexity and uncertainty beyond prior Codex methods.
One key insight: agents do not fit neatly as static elements in state machines. Models improve and address larger issues than initial templates anticipate. Teams now assign goals instead of rigid steps, similar to managers setting outcomes for staff without detailed instructions.
The team states, "The power of models comes from their ability to reason, so give them tools and context and let them cook."
Simple Markdown Drives the System
The Symphony repository centers on a SPEC.md file. This Markdown document outlines the issue and solution. Rather than a full monitoring setup, OpenAI provides a spec for agents to follow. The reference version uses Elixir, chosen for its concurrency features. Codex produced this in one go. To verify, the team generated versions in TypeScript, Go, Rust, Java, and Python.
A WORKFLOW.md details steps like accepting tickets, checking repositories, updating status, linking pull requests, and adding videos. Symphony supplies this to agents. Updates involve editing the file, then directing agents to the revision.
Boundaries and Community Extensions
Symphony suits not all work. Tasks needing ambiguity resolution or judgment stay with developers in direct Codex interactions. It primarily absorbs routine duties, freeing focus for challenging issues.
OpenAI views Symphony as a reference, not a maintained product. Community members produced forks quickly, including one for Anthropic's Claude Code using GitHub Issues. Code and specs reside on GitHub.
This joins other OpenAI agent efforts. In mid-April, ChatGPT gained workspace agents powered by Codex. These automate team processes, have dedicated spaces, connect to Slack, and persist offline.
OpenAI, founded in 2015, develops advanced AI models including the GPT series and code-focused Codex. Linear, established by Karri Saarinen, offers project management for software teams, emphasizing speed and integration.