OpenAI Open-Sources Symphony: An Autonomous Coding Agent Orchestrator

Six months ago, an internal team at ¹ made a controversial call: build a repo with zero human‑written code. Every line had to come from Codex. The experiment worked — until they hit a new bottleneck. Engineers could comfortably manage three to five concurrent Codex sessions before context switching became painful. Beyond that, productivity collapsed.

Their solution is Symphony — an open‑source orchestrator that turns project management boards like Linear into a control plane for coding agents. Instead of humans steering individual Codex sessions, Symphony polls the task board, assigns agents to every open ticket, and runs them autonomously until work lands in the review column.

Symphony is not a product. It is a SPEC.md file paired with a working reference implementation in Elixir/OTP, published under Apache 2.0. OpenAI is explicit that they do not plan to maintain it as a standalone product — it is a blueprint other teams can fork, port, or rebuild, according to OpenAI's announcement.

The core loop is brutally simple. Every open Linear issue maps to a dedicated agent workspace. Symphony polls on a fixed cadence, picks up new tickets with bounded concurrency, and runs a Codex session until that ticket reaches review. If an agent crashes, Symphony restarts it. If a turn times out, it retries. If a new ticket appears, it picks it up.

Developer Chew Loong Nian ran a four‑day experiment and documented the results on Towards AI. He pointed Symphony at 20 real Linear issues from a side‑project repo on a Wednesday morning. By Sunday night, 14 issues had merged pull requests with green CI, three were sitting in code review, and three had been bounced back with the agent explaining why it could not make progress.

"I attended zero standups. I wrote zero prompts. I reviewed 14 PRs in the time it normally takes me to triage a backlog," he wrote. Symphony hit 15,000 GitHub stars in roughly five days after its April 28, 2026 release.

But Chew also noted that the 500% PR increase claim from OpenAI's announcement is both "true and dangerously misleading." The orchestrator works best for well‑scoped, clearly defined tasks. Ambiguous issues produced stalled agents and low‑quality output. "The specific kinds of work where it falls flat on its face" include tasks requiring architectural judgment or cross‑cutting changes.

OpenAI chose Elixir for the reference implementation because of its "excellent primitives for orchestrating and supervising concurrent processes," according to the.¹ The Erlang VM (BEAM) was designed for telecom switches — systems that must manage thousands of concurrent connections, recover from failures gracefully, and never go down. Those same properties map directly to agent orchestration: hundreds of concurrent Codex sessions, each needing isolated state, fault tolerance, and monitoring.

The choice signals something important. Symphony is not a Python library you pip install. It is infrastructure software — the kind where language selection matters for reliability, not convenience. Teams that want to run Symphony will need Elixir expertise or will port the SPEC.md to their own stack, as Tessl noted in their analysis of the release.

Symphony is intentionally not a monolithic supervision system. It is defined as a SPEC.md file that describes the orchestrator's requirements and solution. Teams can use this specification to build their own orchestration layer in any language, as InfoQ reported. The reference implementation on GitHub demonstrates the concepts but is not the only way to implement them.

This "software as a spec" approach flips the traditional model. Instead of shipping a tool and writing docs later, OpenAI shipped the spec first and the tool second. The SPEC.md is the source of truth. The Elixir code is one interpretation of it. Other interpretations — in Python, Go, Rust — are explicitly encouraged.

The shift from interactive coding agents to autonomous orchestration is not just a tooling upgrade — it changes who does what on an engineering team. In the Symphony model, senior engineers spend less time writing code and more time writing excellent issues. The quality of the task description becomes the quality of the output. Vague tickets produce vague PRs. Well‑scoped tickets with clear acceptance criteria produce mergable code.

For teams already using AI coding tools like Cursor, Copilot, or Codex in interactive mode, Symphony represents a different workflow entirely. It is not "faster coding." It is coding while you sleep. The agents run continuously, 24/7, and humans show up to review the results. This shifts the human bottleneck from doing to deciding.

Symphony has clear limitations that early adopters should understand before adopting it. The orchestrator is only as good as the tasks it receives. Architectural decisions, cross‑cutting refactors, and work requiring deep contextual understanding of the codebase remain firmly in human territory. Chew's four‑day experiment found that three out of 20 issues were bounced back because the agent recognized its own limitations — which is arguably the best possible failure mode.

There is also the question of cost. Running dozens of concurrent Codex sessions continuously is not cheap. The 500% PR increase cited by ¹ came from teams with essentially unlimited API budgets. Smaller teams will need to weigh the productivity gains against the inference costs.

Symphony is one piece of a larger shift in how OpenAI thinks about coding agents. The company's earlier configuration engineering blog post laid out the philosophy: treat Codex as a full‑fledged teammate, build agent‑friendly repositories, and invest heavily in automated tests and guardrails. Symphony is the orchestration layer that makes that philosophy scale.

For builders, the message is clear. The future of AI‑assisted development is not a smarter autocomplete. It is autonomous agents that work while you sleep, review while you eat, and ask for help only when they are truly stuck. Symphony is the first major open‑source blueprint for how to build that future. The SPEC.md is available now on GitHub under Apache 2.0.