OpenAI Symphony: ticket-level planning replaces session babysitting, not outcome checks

Primary account: Alex Kotliarskyi, Victor Zhu and Zach Brock, April 27, 2026, plus the public Symphony specification. OpenAI uses Linear tickets as a state machine: one issue can span several PRs or only research, an isolated workspace/agent follows active work, and a human reviews results. Agents produce implementation plans for human approval, derive dependency-aware issue trees, and file out-of-scope improvements as new tasks. PM and design can originate feature work and receive video result packets; CI and rebases can be handled automatically.

Tradeoff: authors report three-to-five interactive sessions were a human-attention ceiling; some internal teams saw +500% landed PRs in the first three weeks with Symphony, but no controlled cost/quality/user-outcome study. Ticket-level automation reduced micromanagement but lost mid-flight nudging and sometimes missed intent; they added tests, screenshots/video and skills to make output legible. Ambiguous/high-judgment work still calls for interactive human engineer. The public spec allows handoff at Human Review and explicitly does not prescribe one sandbox/approval policy. Do not confuse with OpenAI’s different harness account, where human diff review can be optional.

Research gap: The published React/Vite dependency example establishes task-order planning, not a documented customer-to-maintained-feature outcome or human-time ledger. Connect to planning question, integrating question.