OpenAI has launched Symphony, an open-source specification for orchestrating coding brokers round software program duties. It hyperlinks project-management boards to steady agent workflows.
Symphony is designed to show a difficulty tracker resembling Linear into what OpenAI describes as a management aircraft for coding brokers. On this setup, every open activity is assigned to a devoted agent workspace, whereas the system screens progress, restarts stalled brokers and retains work working till the duty is full.
The discharge displays a change in how OpenAI manages software program growth with Codex. Engineers had been utilizing a number of interactive coding periods without delay, however that created a bottleneck as a result of workers needed to maintain switching context between duties, terminals and critiques.
Relatively than centring work on coding periods or particular person pull requests, Symphony is organised round tickets and activity states. That permits brokers to deal with a wider vary of labor, together with investigations and evaluation that won’t end in code adjustments, in addition to bigger growth tasks unfold throughout a number of repositories.
OpenAI has introduced the challenge as a written specification quite than a traditional software program product. The repository is centred on a SPEC.md file that defines the issue and supposed behaviour, whereas a reference implementation is written in Elixir.
The system emerged from OpenAI’s inside engineering workflow. Some groups recorded a 500% enhance in landed pull requests over the primary three weeks of utilizing Symphony.
That enhance got here as engineers spent much less time supervising particular person agent periods. The decrease value of beginning speculative work additionally modified group behaviour, making it simpler to attempt concepts, check refactors and discard weak outcomes with out committing vital workers time.
Symphony additionally widened entry to the event course of. Product managers and designers can file function requests straight into the workflow and obtain a overview packet that features a video walkthrough of the ensuing work within the product.
The way it works
The orchestration mannequin makes use of the duty supervisor as a state machine. Brokers start solely on duties that aren’t blocked, permitting work to proceed in parallel the place dependencies allow.
In a single instance, a React improve was marked as blocked till a migration to Vite had completed. Brokers began the React work solely after that earlier migration was full.
Brokers can even add follow-up duties once they detect points exterior the unique scope, resembling a refactoring alternative or an architectural change. These duties can then be reviewed and scheduled in the identical workflow.
The orchestrator can even help massive monorepos, the place the method of touchdown a pull request could also be gradual or fragile. In these circumstances, Symphony watches steady integration checks, rebases adjustments when wanted, resolves conflicts and retries failed checks.
Commerce-offs
OpenAI mentioned the method has limits. Shifting from direct interactive supervision to ticket-level project reduces engineers’ means to intervene constantly whereas an agent is working, and a few duties nonetheless require hands-on work in interactive Codex periods.
These are typically issues that require stronger judgement or stay extremely ambiguous. Routine implementation work, against this, might be delegated extra readily, leaving engineers to give attention to more durable duties one after the other.
OpenAI additionally mentioned early variations of its agent workflow have been too inflexible as a result of they targeted narrowly on implementing a activity. Over time, the corporate expanded the instruments obtainable to brokers so they may learn overview suggestions, examine CI logs, handle pull requests and report on accomplished or deserted work.
That led to a broader shift from giving brokers mounted transitions to assigning them goals. OpenAI mentioned the change was supposed to make higher use of the reasoning means of newer fashions.
Constructed by brokers
Symphony itself was developed with heavy use of Codex. An early model ran in tmux and polled Linear whereas spawning sub-agents for brand new work, earlier than later iterations have been built-in into an inside repository designed to help agent-led growth.
OpenAI then extracted the idea right into a standalone specification and requested Codex to implement it in Elixir. To refine the design, it additionally used Codex to implement Symphony in TypeScript, Go, Rust, Java and Python to reveal ambiguities within the specification.
The printed model now not relies on OpenAI’s inside repositories or workflows. Alongside the primary specification, the method brokers are anticipated to observe is documented in a WORKFLOW.md file, which units out steps resembling shifting points via statuses, attaching pull requests and including supporting materials.
OpenAI mentioned Symphony additionally depends on Codex App Server, a headless mode that permits programmatic interplay via a JSON-RPC API. That gave the corporate a technique to combine Codex into its workflow with out counting on reside command-line periods.
Though OpenAI has launched the challenge overtly, it doesn’t plan to take care of Symphony as a standalone product. As a substitute, it sees the challenge as a reference implementation for groups that need to construct their very own orchestration methods round coding brokers. OpenAI added that the repository had already handed 15,000 GitHub stars.