Symphony¶
OpenAI's open-sourced orchestration framework for running large numbers of coding agents, released by openai-frontier alongside Ryan Lopopolo's harness-engineering essay. Distributed as a ghost-library: a spec-only reference repo + an Elixir implementation derived by Codex. Users don't consume it as a dependency; a coding agent re-derives the library inside the target codebase from the spec.
Why Elixir¶
Not a taste choice — Codex picked it. GenServer and OTP process supervision "are super amenable to the type of process orchestration" Symphony does: every in-flight task runs as a little daemon driven to completion. Lopopolo explicitly says his own (non-)ability to write Elixir is irrelevant because "no humans in the loop here" — the right tool for the job wins regardless of team language fluency, a subtle but durable consequence of code-is-free.
Six reflection layers¶
Symphony reifies Ryan's mental model of agentic software into six extraction/reflection levels — plus a zero-layer for meta-improvement (can the agent modify its own workflow MD?):
- Policy — what's allowed / disallowed
- Configuration — parameters and knobs
- Coordination — how agents divide work
- Execution — the actual task runs
- Integration — wiring to external systems
- Observability — tracing and introspection
Each layer is something you can tune, distill skills into, and hand to agents to improve. The zero-layer is where skill-distillation lives.