Symphony: AI Coding's Real Bottleneck Is Human Attention

OpenAI's Symphony repo is not just an orchestrator for coding agents; it represents a deliberate shift in how engineering teams interact with AI. The core insight is that the bottleneck has moved from model capability to human cognitive load. Context-switching between multiple interactive coding sessions is unsustainable, leading to errors like issuing an instruction to the wrong thread. Symphony's solution is to elevate the human's role from session supervisor to ticket manager. It does this through a background process that polls a project management tool (Linear, by default) every 30 seconds. When a ticket is in 'to-do,' the system spins up an isolated workspace, launches an agent, and lets the agent work on the deliverable end-to-end. The ticket itself becomes the durable state machine for human-agent communication—agents report plans, logs, and video evidence of testing directly in the ticket. This design is absurdly simple but effective. Crucially, the entire configuration sits inside a single, version-controlled `workflow.md` file in your repo. The top part is YAML that configures the scheduler (which project to monitor, parallelism limits, hooks). The bottom part is a prompt that serves as the agent's standard operating procedure for any ticket. There is no separate config service, admin panel, or UI. Changing the process is a pull request. The real-world viability of this paradigm, however, hinges entirely on a concept the demo calls 'harness engineering': the codebase must be so well-prepared that an agent can complete tasks atomically. The agent needs a bootable system, excellent documentation, and—most critically—self-verifying testing tools that can produce artifacts like video recordings. Without this foundation, Symphony is just a task launcher for unreliable agents. The community is already building implementations in different languages and for different agents, but teams should understand that the orchestration is the easy part; the necessary base investment is in making the codebase truly agent-operable.