Justy: So your issue tracker is just sitting there, collecting tasks. What if every open ticket automatically got a coding agent assigned to it?
Cody: That's Symphony. And honestly, it's clever because it flips the whole model. Instead of you managing five Codex sessions in different tabs, juggling context, you just file a ticket and the system handles it.
Justy: Welcome to Exploring Next, episode 326. Today we're talking about Symphony, an open-source spec from OpenAI that turns your issue tracker into an always-on coding agent orchestrator. Cody, why does this matter right now?
Cody: Because the scaling problem with coding agents isn't the agents themselves — they're fast. It's the human managing them. OpenAI's team hit a wall. Beyond three or four concurrent agent sessions, context-switching tanked productivity. They'd built this team of really capable junior engineers and then spent all their time micromanaging them. That's not sustainable.
Justy: So the insight is, stop treating it like interactive sessions. Treat it like work that gets pulled from a backlog.
Cody: Exactly. They realized the real unit of work isn't a session or a PR — it's a ticket. So they built Symphony to watch Linear and continuously assign agents to open tasks. Each task gets its own agent workspace.
Justy: How does it handle dependencies? Like, if one task blocks another?
Cody: They use a DAG — a directed acyclic graph. Dependencies unfold automatically. And agents can create their own follow-up tasks if they spot issues outside the current scope.
Justy: So one task could generate multiple PRs across different repos?
Cody: Yeah. The ticket is the abstraction layer. Once you decouple work from sessions and PRs, you can orchestrate much bigger units of work.
Justy: And the actual implementation is just a SPEC.md file, right? It's not a framework you download.
Cody: Right. It's a spec — language-agnostic. OpenAI used agents to help implement Symphony itself.
Justy: Let's talk adoption. Who actually uses this? Is it just OpenAI?
Cody: Linear saw a spike in workspace creation after this dropped. But the barrier is real. You need solid CI/CD, good test coverage, decent documentation. OpenAI saw a 500% increase in landed PRs on some teams in the first three weeks.
Justy: That's a huge number. But what breaks? What's the honest trade-off?
Cody: You lose the ability to nudge agents mid-flight. With Symphony, you assign work and review the result. But OpenAI found that useful — failures revealed gaps. They built better tests and clearer definitions of what done looks like. And don't box agents in with rigid state machines. Give them objectives and tools and let them reason.
Justy: That's interesting because it means you're treating the model as smarter than you initially assumed it was.
Cody: Right. Models get better and can solve bigger problems than the box you try to fit them in. The power comes from reasoning, so constrain less and enable more.
Justy: Alright, Build Next. Cody, what's a real way to start experimenting with this?
Cody: Start with the spec itself — it's on GitHub. If you're using Linear, implement a basic version. Watch your issue tracker and assign open tasks to an agent via the API. You don't need the full Symphony yet — just the loop. You could also look at existing orchestration libraries like Langchain or CrewAI, which have agent coordination primitives.
Justy: That's Exploring Next, episode 326. Symphony is a shift from supervising agents to orchestrating their work. Thanks for digging in, Cody.
Cody: Thanks, Justy.