Symphony: Orchestrate Coding Agents via Tickets, Not Sessions
OpenAI's Symphony automates coding agents at ticket level using Linear as a state machine; run once, it polls every 30s, spins isolated workspaces, and follows workflow.md for end-to-end task completion without human session management.
Ticket-Level Oversight Unlocks Scalable Agent Management
Current coding agent workflows overload humans with 2-3 parallel sessions, leading to context-switching errors and cognitive limits that cap output below model potential. Symphony reframes this by elevating humans to manage tickets (e.g., in Linear) instead of sessions. A background scheduler polls your Linear board every 30 seconds for 'to-do' tickets, creates isolated workspaces per ticket, launches agents, and updates ticket status (to 'in progress', 'human review', 'merging'). Agents report directly to tickets with plans, checklists, and video proofs, mimicking how engineering leaders oversee thousands of tasks via outcomes, not PRs. This scales beyond 3 sessions since humans intervene only for reviews or merges, not monitoring.
Workflow.md Encodes Scheduler Config and Agent SOP in One File
The single workflow.md file in your repo drives everything via YAML frontmatter (project slug, API keys, poll interval, parallel agents, post-workspace hooks, agent settings like CodeX config) and markdown body as the agent's persistent prompt. It details SOP: task planning, validation, 'done' criteria, human outreach triggers. Version-controlled via PRs, it eliminates separate UIs/config services; update it to onboard new agent capabilities. Flexible beyond Linear/CodeX—adapt via spec.md to any ticket tool/language (e.g., community ports to Python, TUI, Cloud Code). No admin overhead; same file controls scheduler and agent behavior.
Codebase Harness Enables Atomic End-to-End Completion
Agents fail without a 'harness': bootable env (scripts auto-setup), docs index (agent.md/codex.md), and self-verification. Add Playwright CRI for browser testing with video recording (video.start/stop commands capture MP4/WebM, overlay annotations/chapters, upload to Linear for proof). Include skills for server start, Linear API ops (status updates, video uploads), production logs (e.g., Grafana fetch), debugging. Predefine scripts for complex boots. These make agents autonomous: implement, test E2E, verify via video, report—without human babysitting. Copy-paste from AI Build Club repos; useful even sans Symphony.
Zero-to-Running Setup Delivers Immediate Workflow
Clone Symphony (reuse OpenAI's Elixir impl or agent-build Python via spec.md). Install Linear: create project (note slug), get API key (LINAPI_KEY env). Define statuses: 'to-do' → agent pickup; 'human review' post-completion; 'merging' → auto-PR. Agent-generate workflow.md pointing to your repo. Run symphony path/to/workflow.md --dangerously-skip-guardrails for daemon mode. Create Kanban view; drop ticket to 'to-do' (e.g., 'Change hero copy'); watch agent plan, execute in isolated workspace, update status, upload video. Review video/PR, approve merge. Dashboards track sessions; scales via parallelism config.