We just published a blog post about Symphony – an agent orchestrator I built for my team at OpenAI. Here I wanted to share a few behind-the-scenes bits that didn’t make the cut. Earlier this year I joined a team that had gone all-in on Codex. And it worked! It worked because we designed the software in a way that maximizes the benefits and keeps technical debt under control. Now we were constrained on human attention. Even the smartest engineers lose track of parallel Codex sessions after working on half a dozen tasks for hours. Symphony started as a way to make my teammates happier and break the ceiling of the 5 Codex sessions per person. Basic Idea In short, Symphony is an orchestrator – it takes a Linear board, and for each open unblocked task it makes sure an agent is running. Think Kubernetes for agents where the task board is the control plane. On top of that it allows us to define success criteria and shift from steering agents to reviewing their “proof of work” packets.…
No comments yet. Log in to reply on the Fediverse. Comments will appear here.