OpenAI Frontier — Symphony and the "Ghost Repo"
Source: Ryan Lopopolo (OpenAI Frontier team) interview on Latent Space, 2026-04 A deep dive into how OpenAI's Frontier team built Symphony — a codebase with 1 million lines of code, zero lines written by humans, merged without human review.
The Frontier team
- OpenAI's newest team, focused on helping enterprises safely deploy agents at scale
- Became OpenAI's largest Codex user
- Started with 3 people, spent 5 months on extreme experiments
- Final vision: a "distribution spec" system that plugs into enterprise IM, security tools, and workflow tools
The Ghost Repo: Symphony
Symphony doesn't contain actual code. Instead, it contains all the context, specs, and workflows needed for AI agents to autonomously generate 1M lines of code.
Core insight from Ryan:
"When AI fails, don't immediately try to improve the prompt. Ask: what capability, context, or structure is missing?"
This mindset shift increased their development speed by 10x.
The 7-layer architecture
Symphony's architecture has 7 layers (detailed in the podcast), designed to provide the complete scaffold for agentic code generation.
Key engineering disciplines
The one-minute build rule
To maximize inner-loop speed, the team enforces a strict "one-minute build discipline". If a build takes longer than one minute, they refactor until the agent's feedback loop is short enough.
This produces code that is:
- Highly modular
- Observable
- Token-efficient
MCP is dead
Ryan is pessimistic about MCP:
"It forces massive token injection into context, breaks auto-compaction, and agents may forget how to use the tools."
更新：Codex 开源 Symphony（2026-04-28）
OpenAI 通过 @OpenAIDevs 开源了另一个 Symphony——这是面向公众的轻量级 Codex 编排层，与 Frontier 团队的内部 Ghost Repo 是不同项目：
| 维度 | Frontier Symphony（Ghost Repo） | Codex 开源 Symphony |
|---|---|---|
| 定位 | 企业级 agent 分发基础设施 | issue tracker → agent → PR 工作流 |
| 规模 | 1M LOC，零人工编写 | 轻量级开源库 |
| 使用场景 | 大型企业安全部署 | 开源社区、中小团队 |
| 与 MCP 关系 | Ryan Lopopolo 认为 MCP 已死 | 作为 Codex skill 使用，与 MCP 生态共存 |
核心工作流：open issue → agent 生成代码 → PR → human review。标志着 OpenAI 正在构建更完整的 agent 基础设施栈，而不仅仅是提供模型 API。
Redefinitions for the agent era
Ryan argues that many software concepts need redefinition:
| Old concept | Agent-native redefinition |
|---|---|
| Software dependencies | May disappear — medium-complexity software (thousands of lines) can be rewritten by AI and internalized by models |
| Bug | "Code written by an agent that is inconsistent with a not-yet-written non-functional requirement" |
| MVC pattern | Model-View-Claw — where Claw = harness |
| Software readability | Must be agent-readable first; implicit context kills agent effectiveness |
Culture and scale
- Humor as AGI test: The team teaches agents company culture, including generating memes and interacting with Slack. "Humor is part of AGI."
- Token consumption: The team burns 1 billion tokens per day (~$2,000–$3,000). Ryan: "If you're not doing this, you're negligent. Stingy with tokens = stingy with efficiency."
- Codex growth: Codex app weekly active users surpassed 2M, growing at 25% per week.
- Geographic expansion: OpenAI is expanding beyond SF to Seattle, NYC, London. Ryan was one of the first engineering hires for the Seattle (Bellevue) office.
Counterpoints & Gaps
- The "zero human review" claim depends on extremely mature specs and automated validation; most organizations lack this foundation.
- "MCP is dead" is a strong claim from a team with full control over their toolchain; for heterogeneous tool ecosystems, standards like MCP may still provide value.
- The 1B tokens/day figure assumes enterprise budgets; individual builders and small teams cannot replicate this burn rate.
Related
- harness-engineering/overview — Harness engineering overview
- harness-engineering/fat-skills-fat-code-thin-harness — Garry Tan's thin harness philosophy
- harness-engineering/ai-native-dev-tencent-openspec — Tencent's AI-native dev practices
- product-trends/headless-mode-survival-requirement — Headless software trends