Home
Zero People Coding: OpenAI Engineers Reveal AI Distillation for Agents as the Ultimate Teammate, 10M Lines of Code

OpenAI engineer Ryan Lopopolo recently shared an extensive article, "Harness Engineering," detailing how internal teams heavily rely on Codex. This experiment gave rise to a "ghost library" named Symphony and established a new engineering paradigm.
Core Experiment: Embracing a "No Human Coding" Constraint
Ryan imposed a dramatic constraint at the experiment's outset: never manually write any code.
From Frustration to Breakthrough: Initially, the model struggled with complex logic, making development ten times slower than human output.
Driving Tool Evolution: To let the Agent work independently, the team built a highly refined Harness (Development Console) system.
Exponential Gains in Efficiency: As models like GPT-5.4 improved, the system's efficiency eventually surpassed any single engineering team.
Key Tactics: Keeping Build Time Under "One Minute"
Inside OpenAI , the Agent's efficiency hinges on extremely fast feedback:
Inner Loop: The team iteratively rebuilt the build system (from Makefile to Bazel to Nx), keeping build times strictly under 1 minute.
Logical Decomposition: When builds slowed, the Agent automatically split the build graph into smaller pieces until complexity dropped below the threshold, preserving the Agent's focus.
Paradigm Shift: Humans Become the Bottleneck in PR Review
When code could be easily parallelized, the true scarce resource was human attention.
Post-Merge Review: The team automated pre-merge reviews, shifting most manual review to after code was merged for quality assurance and knowledge capture.
Observability as Empowerment: Engineers now focus on equipping the Agent with Traces (Tracing) and observability tools, granting it self-healing abilities, rather than fixing bugs themselves.
Experience Distillation: Senior engineers codify their tacit knowledge into Skill documents and tests, embedding it into the system context.
Future of the Industry: Ghost Libraries and Dependency Internalization
Ryan offered a provocative vision: as token costs near zero, software dependencies could vanish.
Ghost Libraries: Developers only need to define a high-fidelity spec, letting the Agent reconstruct and implement it locally.
Dependency Internalization: For medium-to-low-complexity third-party libraries, the Agent can inline them into the repository, stripping extraneous code and creating a fully self-contained tech stack.
Conclusion: From Copilot to Autonomous Teammate
According to OpenAI's Frontier Team, future software development will optimize for Agent readability rather than human habits, restructuring the entire codebase accordingly. When Agents manage the full lifecycle—from writing code and CI deployment to production monitoring—human engineers will shift entirely to being the "shepherd" of the system architecture.
Related article
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber
Related Special Topic Recommendations
Comments (0)
0/500

OpenAI engineer Ryan Lopopolo recently shared an extensive article, "Harness Engineering," detailing how internal teams heavily rely on Codex. This experiment gave rise to a "ghost library" named Symphony and established a new engineering paradigm.
Core Experiment: Embracing a "No Human Coding" Constraint
Ryan imposed a dramatic constraint at the experiment's outset: never manually write any code.
From Frustration to Breakthrough: Initially, the model struggled with complex logic, making development ten times slower than human output.
Driving Tool Evolution: To let the Agent work independently, the team built a highly refined
Exponential Gains in Efficiency: As models like
Key Tactics: Keeping Build Time Under "One Minute"
Inside
Inner Loop: The team iteratively rebuilt the build system (from Makefile to Bazel to Nx), keeping build times strictly under 1 minute.
Logical Decomposition: When builds slowed, the Agent automatically split the build graph into smaller pieces until complexity dropped below the threshold, preserving the Agent's focus.
Paradigm Shift: Humans Become the Bottleneck in PR Review
When code could be easily parallelized, the true scarce resource was human attention.
Post-Merge Review: The team automated pre-merge reviews, shifting most manual review to after code was merged for quality assurance and knowledge capture.
Observability as Empowerment: Engineers now focus on equipping the Agent with
Experience Distillation: Senior engineers codify their tacit knowledge into Skill documents and tests, embedding it into the system context.
Future of the Industry: Ghost Libraries and Dependency Internalization
Ryan offered a provocative vision: as token costs near zero, software dependencies could vanish.
Ghost Libraries: Developers only need to define a high-fidelity spec, letting the Agent reconstruct and implement it locally.
Dependency Internalization: For medium-to-low-complexity third-party libraries, the Agent can inline them into the repository, stripping extraneous code and creating a fully self-contained tech stack.
Conclusion: From Copilot to Autonomous Teammate
According to OpenAI's Frontier Team, future software development will optimize for Agent readability rather than human habits, restructuring the entire codebase accordingly. When Agents manage the full lifecycle—from writing code and CI deployment to production monitoring—human engineers will shift entirely to being the "shepherd" of the system architecture.
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber











