What it does

Symphony watches a project management board (like Linear), spawns agents to pick up tasks, and doesn’t stop until the code lands. The agents produce proof of work—CI results, PR reviews, complexity analysis, even walkthrough videos—and merge only when accepted. It’s pitched as the next step after “harness engineering”: stop managing agents, start managing work.

The interesting bit

The project is almost aggressively self-referential. OpenAI published a spec and an Elixir reference implementation, but the first recommended setup path is literally: “tell your favorite coding agent to build Symphony.” It’s a framework for autonomous agents that you can have an agent build for you.

Key highlights

• Monitors external task boards and dispatches agents automatically
• Agents provide structured proof of work before requesting merge
• Built in Elixir, which handles concurrency and fault tolerance well for long-running autonomous processes
• Ships as a low-key engineering preview—OpenAI’s own warning, not mine
• Reference implementation lives in elixir/; the spec is language-agnostic

Caveats

• Explicitly marked for “testing in trusted environments” only
• Requires prior adoption of “harness engineering” practices; unclear how much upfront investment that demands
• 24,985 stars but no topics, no detailed docs beyond setup—very much a preview

Verdict

Worth watching if you’re already running coding agents at scale and want to automate the orchestration layer. Skip it if you’re looking for a polished product; this is a sketch of where OpenAI thinks agent management is heading.