Symphony — codex orchestrator from openai

No worries, no worries. Soon they’ll add multi-agents and decouple from linear, just like in flowai-workflow :) Although running different agents and ides for different stages still won’t work for them.

Even isolated agents with self-checks can already close a fairly large percentage of tasks. Especially boring stuff like refactoring, documentation, tech-debt cleanup, and so on. The downside is the accumulation of unknowns in your own code, but that’s covered by:

• detailed task descriptions
• the agent’s ability to stop or request human help
• strict review
• quality gates

https://openai.com/index/open-source-codex-orchestration-symphony/

TL;DR: Symphony is a minimalist orchestrator for code agents (codex) that pulls tasks from a tracker (linear), creates a separate working directory for each, and guarantees that an agent is constantly working on every active task until it reaches the desired status (e.g., human review).

Core idea

The openai team got tired of manually running codex sessions for simple tasks and jumping between them. They shifted focus from “sessions and PRs” to “tasks in the tracker” and built Symphony — a layer that reads tasks from linear and starts/restarts agents on its own.

How it works at a high level

Symphony turns the task board into a state machine. Any open task in an active status gets its own agent and workspace (a directory on disk). The orchestrator polls linear in a loop, distributes tasks based on priorities, blockers, and parallelism limits, watches for crashes/hangs, and retries with exponential backoff. Agents can create new tasks themselves, break work into sub-tasks, and build a dag of dependencies.