오픈AI, 인간 개입 최소화하는 AI 에이전트 자율 관리 시스템 공개

OpenAI says human attention is the bottleneck, so it built a system to let agents manage themselves Jonathan Kemper View the LinkedIn Profile of Jonathan Kemper May 4, 2026 Nano Banana Pro prompted by THE DECODER Key Points OpenAI has released Symphony, an open-source specification that transforms task trackers like Linear into a control center for AI agents, allowing them to autonomously process open tickets while developers simply review the results. Symphony addresses the challenge of managing multiple AI sessions simultaneously by automatically assigning active tickets to agents, which can also create new tickets for follow-up tasks on their own when needed. The agents are guided through simple markdown files with clear objectives, and while OpenAI provides Symphony only as a reference implementation, the community is already adapting the system for other models and platforms. Ask about this article… Search OpenAI has released Symphony, an open-source spec with a reference implementation that turns task trackers like Linear into a command center for Codex agents. Instead of developers juggling multiple sessions, the agents grab open tickets themselves, leaving humans to review the results. Some internal teams saw merged pull requests jump sixfold in the first three weeks, according to OpenAI. Linear founder Karri Saarinen also noticed a spike in new workspaces in his project planning tool after the release. With Symphony, each open ticket gets its own Codex agent and dedicated workspace that runs until the task is done, turning the task board into the place where work actually gets dispatched. Ad When human attention becomes the bottleneck Before Symphony, OpenAI developers ran several Codex sessions in parallel, handed out tasks, and chased down progress on each one, according to OpenAI. In practice, running more than three to five sessions at once was nearly impossible without constant context-switching tanking productivity. Ad DEC_D_Incontent-1 "The agents were fast, but we had a system bottleneck: human attention," the developers write. They had built a team of junior developers, only to stick their human coworkers with the micromanagement. That's what sparked the idea of flipping the workflow. Instead of supervising sessions, agents would pull their own work straight from the tracker. Linear as a state machine Symphony uses Linear as a state machine. Tickets move through statuses like "Todo," "In Progress," "Review," and "Merging." The system watches the board and makes sure every active ticket has an agent assigned to it. If an agent crashes or stalls, Symphony spins it back up. Ad Only unblocked tickets get picked up, which lets a task tree run in parallel. The team points to a React upgrade that only kicked off after an upstream migration to Vite. Tickets can be much bigger than a single code change. Some spawn several pull requests across different repos, while others are purely research or analysis tasks with no code at all. If agents spot issues outside their current ticket along the way, like performance problems or refactoring opportunities, they file new tickets on their own. OpenAI says product managers and designers can also submit feature requests directly and get back a review package with a video walkthrough without ever checking out the repo. Ad DEC_D_Incontent-2 One lesson from building Symphony: agents are hard to treat as fixed nodes in a state machine. The models keep getting better and can tackle bigger problems than the template plans for. The team now prefers to hand agents goals rather than strict processes, the way a manager gives an employee a result to deliver, not a step-by-step playbook. Ad "The power of models comes from their ability to reason, so give them tools and context and let them cook," the team says. A Markdown file at its core Open the Symphony repo and you'll mostly find a SPEC.md. The Markdown file lays out the problem and the desired solution. Instead of building a complex monitoring system, OpenAI ships a spec that agents can implement themselves. The reference implementation is written in Elixir, a language with solid tools for concurrent processes. OpenAI says Codex built this implementation in one shot. To stress-test the spec, the Codex team also had it implemented in TypeScript, Go, Rust, Java, and Python. The development workflow—accept ticket, check out repo, set status, attach PR, attach video—lives in a WORKFLOW.md that Symphony hands to agents as a guide. To change the process, edit the file, and Symphony points the agents to the new version. Limits, forks, and what's next Not every task is a fit for Symphony. Ambiguous problems or work that calls for judgment still get handled directly by developers in interactive Codex sessions, OpenAI says. Symphony is mainly there to soak up routine work so people can focus on one hard problem at a time. OpenAI doesn't plan to maintain Symphony as a standalone product and sees it more as a reference. The community has already shipped forks, including one for Anthropic's Claude Code with GitHub Issues . Code and specs are available on GitHub . Symphony is one of several agent projects at OpenAI. In mid-April, the company rolled out workspace agents in ChatGPT , also powered by Codex, aimed at automating complex team workflows. They have their own workspace, plug into Slack, and keep running even when the user is offline. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now Source: OpenAI