From supervising agents to supervising outcomes
One of the practical limits of AI coding tools has not been model capability alone. It has been management overhead. Even when agents can write code, humans often still spend their time opening sessions, assigning tasks, tracking progress, and restarting work when a run stalls. OpenAI’s newly released Symphony specification is designed to reduce that coordination burden by changing who manages the queue.
According to reporting from The Decoder, Symphony is an open-source specification with a reference implementation that turns a task tracker such as Linear into a control system for AI agents. Instead of developers distributing tickets manually across multiple sessions, agents can pull eligible work directly from the board, process it in dedicated workspaces, and return results for human review.
The bottleneck Symphony is trying to solve
The core argument behind the system is strikingly simple: if agents are fast but humans still have to micromanage them, then human attention becomes the throughput constraint. The report says OpenAI developers found it difficult to manage more than roughly three to five simultaneous Codex sessions without losing efficiency to context switching. In that setup, the people were acting less like engineers and more like dispatchers.
Symphony flips that arrangement. The tracker becomes a state machine, with statuses such as Todo, In Progress, Review, and Merging. The system watches those states, ensures each active ticket has an assigned agent, and can restart an agent if it crashes or stalls. Only unblocked tickets are picked up, allowing dependency trees to proceed in parallel where possible.
