OpenAI Symphony: 6x More Merged PRs When Agents Self-Orchestrate
Agentic AIOpenAI’s Symphony converts a task board into an autonomous coding fleet — and some internal teams saw the number of merged pull requests increase sixfold within three weeks of deploying it.
Context
Deploying AI coding agents has become straightforward. Managing them at scale has not. As teams grow from one agent to five to dozens running in parallel, the cognitive overhead of supervising sessions, restarting failures, and directing work starts consuming the productivity gains the agents were supposed to deliver. The engineer becomes a supervisor rather than an architect.
Symphony’s timing fits a pattern visible in Stack Archive’s coverage of agentic coding: Cursor’s parallel fleet model addressed agent concurrency, earlier Copilot integrations addressed SDLC tooling, and now Symphony addresses the coordination and persistence layer sitting above all of it.
The Detail
Symphony is an open-source specification — not a product — that defines how coding agents should interact with a task tracking system. The reference implementation uses Linear as the control plane: every open ticket becomes a task assigned to an agent, which receives its own isolated workspace. The agent checks out the repository, works the ticket to a pull request, updates ticket status along the way, and moves on. If it crashes, Symphony restarts it. If new tickets appear, Symphony picks them up automatically.
The architecture centers on two files: a SPEC.md that describes the overall system, and a WORKFLOW.md that gives agents step-by-step instructions for processing a task. The reference implementation is in Elixir — chosen for its concurrency model — but TypeScript, Go, Rust, Python, and Java versions now exist from community contributors.
Two properties distinguish Symphony from simply running agents manually. First, agents can create their own work: when an agent notices a performance issue or refactoring opportunity during implementation, it can open a new ticket for it, which Symphony then picks up. Second, Linear’s status transitions — Todo, In Progress, Review, Merging — function as the coordination layer between agents and human reviewers. Humans don’t direct work; they review completed work.
The sixfold increase in merged PRs at some OpenAI teams reflects the upstream effect: fewer bottlenecks from crashes, no tickets sitting unattended, and human attention focused on review rather than direction. According to OpenAI’s Symphony announcement, v1.1.0 now supports the Kata CLI as an alternative agent runtime, which means Claude Code, Gemini, and other models can run inside the same orchestration framework.
The Pattern
Stack Archive has covered agentic coding as an engineering discipline in four posts over the past 90 days — agentic SDLC patterns, Cursor’s parallel fleet architecture, coding agent component design, and CI/CD automation for agentic repos. Symphony is the first complete orchestration-layer specification from a major lab, making it a reference point for any team assembling this stack.
The So What
Symphony is a spec, not a SaaS — any team can read it, adapt it, and run it against their own task board. The practical question is whether the underlying model — agents as workers pulling from a queue rather than tools requiring direct prompts — fits how your team wants to operate. For teams already using Linear and deploying Codex or Claude Code agents, the migration path is direct and the spec is open. The 6x PR result is an internal benchmark, not a guarantee — but the architecture that produced it is now available for any team to test.
Photo by Daniil Komov on Pexels
Source: OpenAI
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →