OpenAI Symphony is an open-source, daemon-based framework for orchestrating autonomous AI coding agents at scale — converting issue tracker tickets into completed pull requests without developer supervision. Released March 5, 2026 as an engineering preview (Apache-2.0), it introduces two key primitives: a policy-as-code file (WORKFLOW.md) that version-controls agent behavior alongside application code, and isolated implementation runs powered by Elixir/BEAM for fault-tolerant concurrent execution. 13K+ GitHub stars in three weeks signals strong practitioner interest.
Architecture Deep Dive → OpenAI Symphony Architecture Breakdown — seven-layer architecture, WORKFLOW.md policy-as-code with Jinja2 templating, BEAM/OTP concurrency model, workspace isolation, proof-of-work verification, and comparison with Devin, Open-SWE, and Codex.
Why It's in Assess
Symphony introduces a genuinely new orchestration model for coding agents: instead of prompting agents one-at-a-time, teams manage a kanban board and Symphony runs agents continuously as work appears. This mirrors how platform teams manage CI/CD pipelines — define the policy, let the system execute.
The architectural choices are sound: BEAM/OTP provides battle-tested fault tolerance for running hundreds of concurrent isolated workspaces; WORKFLOW.md treats agent prompts as versioned code rather than ad-hoc strings; SPEC.md provides a language-agnostic specification enabling community re-implementations.
Assess rather than Trial because:
- OpenAI explicitly flags it as engineering preview — not recommended for production use
- Integration is limited to Linear only (no GitHub Issues, Jira, or GitLab)
- Requires significant harness engineering investment: hermetic test suites, structured repo organization, clean CI/CD — teams without this foundation can't use it effectively
- No documented production deployments yet
- The Elixir reference implementation is the only maintained implementation at this time
When Symphony is the right direction: Teams running OpenAI Codex at scale, with mature CI/CD and hermetic test suites, who want to eliminate agent supervision toil. The shift from "supervising agents" to "managing work" is real leverage once the prerequisites are in place.
When to look elsewhere: Teams without mature test infrastructure (Symphony amplifies what exists; it doesn't compensate for what doesn't), teams on GitHub Issues or Jira (not yet supported), teams wanting production-ready tooling today (use Devin or Open-SWE instead).
Core Concepts
WORKFLOW.md — policy-as-code: Lives in the repo root, configures the entire agent behavior in YAML frontmatter. Defines which Linear project to poll, the state machine (Todo → In Progress → Merging → Done/Rework), polling intervals (default 5s), concurrency limits (max 10 agents), per-lifecycle hooks (after_create, before_run, after_run), and Jinja2 templates for dynamic agent prompts. Agent behavior changes are code-reviewed like any other change.
Implementation runs: Each issue gets a deterministic, isolated filesystem workspace. Same issue + same agent configuration = reproducible output. The BEAM runtime manages hundreds of concurrent runs via OTP supervision trees — a worker failure is isolated, never cascading.
Proof-of-work verification: Before a PR lands, Symphony requires CI status passing, unit test results, and automated PR review feedback. Agents that can't meet the quality gate retry or surface to the Rework state rather than landing broken code.
Harness engineering requirement: Symphony is explicit that it amplifies existing engineering infrastructure — it cannot substitute for it. Teams need hermetic tests (no external calls), machine-legible repo structure, and per-layer linting before Symphony adds value.
Key Characteristics
| Property | Value |
|---|---|
| Interface | Daemon service, Phoenix LiveView dashboard |
| Provider | OpenAI |
| License | Apache-2.0 |
| Pricing | Free (open source); requires OpenAI API for Codex execution |
| Language | Elixir/OTP (reference); language-agnostic SPEC.md |
| Issue tracker | Linear (v1); specification supports extensibility |
| Underlying model | OpenAI Codex (GPT-5.4) |
| Status | Engineering preview (not production-recommended) |
| GitHub | openai/symphony |
| Website | github.com/openai/symphony |
| Docs | SPEC.md |