Bypass Human Bottlenecks with Autonomous Task Orchestration

OpenAI built Symphony to address engineers supervising only 3-5 Codex CLI sessions concurrently before context-switching hurt productivity. Instead of manual oversight, Symphony integrates with issue trackers like Linear: agents poll for tasks in 'to-do' states, claim them by updating status, create isolated workspaces matching issue IDs, run Codex to complete work, then mark issues 'done' with comments linking outputs. This scales agent teams without human managers—agents handle routine coding autonomously, flagging humans only for reviews. In demos, it builds a TypeScript Bun 'Hello World' app from scratch, validating GraphQL errors via Codex self-fixes.

Trade-offs: Barebones by design, lacking built-in repo cloning or PR creation, so not production-ready out-of-box. But this enables customization: users add pre/post-hooks in workflow YAML. For example, a 'create after' hook clones a repo to the workspace and checks out a new branch; a 'run after' hook stages changes, commits, pushes, and opens a PR with issue links—transforming Symphony into a full CI/CD agent for apps like film emulation tools.

Prompt-Driven Installation Builds Ownership

Skip cloning Elixir repo (traditional Option 2: clone, build, run via workflow). Use Option 1: Feed agents OpenAI's 2000+ line spec prompt detailing Symphony's architecture. Agents recreate it in any language—Python (easy for LLMs), Go (with Charm CLI), or Claude SDK variants. This genius chaos ensures no two installs match: users own their version, fixing bugs and adding features themselves, reducing OpenAI maintenance. Requires Linear personal API key in YAML workflow (active states like 'to-do', workspace root, Codex command). Tools like UV run Python versions; access OpenAI's YAML from repo.

Why it works: Forces deep understanding, customizes to needs (e.g., swap Linear/Codex). Downside: Chaos for support, but fosters community forks.

Symphony as Customizable Base vs. Polished Alternatives

Position Symphony as 'Pi harness'—minimal orchestrator for exact needs, unlike feature-rich MultiOn (easier setup, multi-agents, scheduling) or Conductor ('Claw Code'). Ideal if using Linear/Codex: leverage existing MCP tools/plugins. Vision: Central agent hub for 'AI dev teams'—shared workflows/tools across tasks, visibility into others' prompts/features to spot conflicts. Extend via Codex for better PR descriptions/links. Not for non-Linear/Codex users yet, but proves agent orchestration scales coding without humans.