Symphony — OpenAI’s Orchestration Spec for Coding Agents

Why this is in the vault

Symphony is OpenAI’s spec for the orchestration layer above coding agents — exactly the layer RDCO has built incrementally over the last 6 weeks. Worth filing for cross-reference + adopting two specific invariants we don’t yet enforce. Don’t port the runtime.

What Symphony is

Apache-2.0 spec + experimental Elixir reference implementation (~20.4k stars, low-key engineering preview at ~12 commits on main). Defines a poll-loop orchestrator that:

Pulls open issues from a tracker (Linear in the reference impl)
Spawns isolated Codex subprocesses, one per issue, in sanitized per-issue workspaces
Reconciles state via the tracker’s own status fields — orchestrator is read-only on the tracker
Crash-recovers via re-polling (no hidden in-memory state)

Not a CLI, not a library, not a managed runtime — a protocol with a sample implementation.

Three sharpest claims

“Manage work, not agents.” Success is defined as workflow-state handoff (e.g. issue moved to “Human Review”), NOT tracker Done. Same thesis as RDCO’s /check-board → critical-component loop.
Orchestrator is read-only on the tracker. State writes happen via the agent’s own tool calls inside its workspace. Clean separation between orchestrator and worker.
Three hard safety invariants treated as load-bearing (not nice-to-have):
- Workspace containment — every command’s cwd must equal the per-issue workspace path
- Sanitized directory names — no .. or shell injection from issue titles
- Crash-recoverable via re-polling — no in-memory state survives orchestrator restart

Mapping against Ray Data Co

RDCO already runs this pattern. The 6-week incremental build of skills + cron + Notion ended up at the same shape Symphony codifies:

Symphony concept	RDCO equivalent
Tracker (Linear)	Notion task board (`collection://ae46c40f-421d-4c94-924b-0a2b13de0756`)
Poll loop	`/check-board` cron (every hour)
Issue-spawned workspace	Sub-agent `Agent` invocations + skill-scoped state files
Codex subprocess per issue	General-purpose sub-agent fan-out (per-newsletter, per-video, per-batch)
Concurrency manager	Skill-spec-defined caps (e.g. /process-newsletter watch caps at 3 concurrent)
Channels as work-status surface	iMessage + Discord (the channels-as-claw pattern, 2026-05-01)

Two borrowable invariants we don’t yet enforce:

Workspace containment per task. Today, sub-agents inherit the parent process’s cwd. We don’t isolate per-task. Low risk in practice (skills generally don’t shell out to filesystem-mutating commands) but Symphony treats this as load-bearing. Worth a Notion task to add per-task cwd-isolation guards to skills that do shell out (build-project, build-landing-page, blender, anything that writes outside the vault).
Explicit “handoff state ≠ Done” Notion field. Today the Notion task board has a Status field (To Do / In Progress / Blocked / Done / Archived). Symphony’s framing argues for an explicit state for “orchestrator-handoff-to-human” distinct from “complete.” The “Blocked” state covers some of this but conflates founder-needed-input vs. code-gate-failed vs. dependency-missing. A dedicated state — e.g. Awaiting Founder Review — would tighten the loop.

Not borrowable: the runtime itself. Elixir + Linear + Codex stack is wrong for us. Stay on Claude Code + Notion + Anthropic + sub-agent fan-out.

Comparison to other agent-architecture filings (today’s cluster)

OpenAI Symphony joins the agent-deployer / orchestration evidence cluster:

Source	Layer	Stack
2026-01-09-trevin-chow-agent-orchestration-thesis	Thesis	Anthropic/Codex
2026-05-01-trevin-compound-engineering-v3-4	Plugin / skill marketplace	Claude Code
2026-05-01-three-primitives-claude-code-life-os	Personal life-OS	Claude Code
2026-05-01-mubbu-personal-brand-five-agent-system	Agency content production	Claude/multi-model
2026-05-01-ann-miura-ko-six-levels-ai-pilled-organizations	Org-design framework	Vendor-neutral
THIS PIECE — Symphony	Orchestration spec for coding agents	OpenAI/Codex/Linear

Two ecosystems (Anthropic/Compound-Engineering and OpenAI/Symphony) converging on the same primitives: tracker → poll → workspace-isolated worker → handoff state. That convergence is itself a signal — the architecture is hardening into industry standard.

Open follow-ups

Workspace containment audit — survey our skills for any that shell out to filesystem-mutating commands without explicit cwd isolation. Queue as Notion task if any are found.
Notion handoff-state field — propose adding an Awaiting Founder Review status (or splitting Blocked into Blocked-Founder-Decision vs Blocked-Dependency) to tighten the orchestrator/handoff distinction.
Watch Symphony’s evolution — 12 commits as of today, low-key preview. If OpenAI ships a more opinionated runtime or pulls Linear-coupling apart, worth re-evaluating.

2026-04-11-garry-tan-thin-harness-fat-skills — Tan’s “thin harness, fat skills” framing the harness-engineering term derives from
2026-05-01-trevin-compound-engineering-v3-4 — Anthropic-side parallel
2026-01-09-trevin-chow-agent-orchestration-thesis — orchestration thesis
2026-05-01-ann-miura-ko-six-levels-ai-pilled-organizations — L4-L5 framework Symphony patterns map to
2026-04-30-rdco-thesis-targeting-systems-feedback-loops — RDCO targeting-systems thesis
2026-04-30-rdco-bet-architecture-playbook — bet architecture