OpenAI engineers said their team used AI agents to produce roughly 1 million lines of code in five months, covering application logic, infrastructure, tools, documentation, and internal developer tooling, while humans wrote zero lines of code by hand. The method behind that workflow is called Harness Engineering. The idea is not to improve the model itself, but to build the surrounding system so agents can code, test, open pull requests, and correct failures with much less direct human intervention.
The numbers OpenAI attached to the experiment
According to the published account, the project began in August 2025. The team grew from 3 people to 7, merged around 1,500 pull requests, and averaged 3.5 PRs per engineer per day. A single AI agent could work autonomously for more than six hours at a time, and the resulting product was already being used daily by hundreds of internal users.
One rule stood out. Humans were not supposed to write code directly at all. That forced the team to invest in agent infrastructure, context management, and workflow design instead of relying on manual fixes whenever the system broke down.
What Harness Engineering is meant to solve
OpenAI described Harness Engineering as the discipline of designing infrastructure, constraints, and feedback loops so AI agents can operate reliably at scale. The “harness” metaphor is central: the model is the powerful but unpredictable horse, the harness is the control system, and the human sets direction while the agent executes.
The framework centers on four functions: constraining, informing, verifying, and correcting. In practice, that means defining boundaries, supplying the right context, checking whether the output is valid, and creating mechanisms that let the system recover from mistakes. The argument is simple: model quality is only part of the equation; the surrounding system decides how dependable the agent becomes.
Three pillars behind the setup
The first pillar is Context Engineering. OpenAI’s principle is that anything an agent needs to know must live inside the repository, which serves as the single source of truth. Instead of turning AGENTS.md into a giant manual, the team treats it like an index that points to structured files such as architecture guides, coding conventions, API references, and execution plans. Static context includes architectural rules, API contracts, and style guides. Dynamic context includes logs, metrics, traces, CI/CD status, and test results.
The second pillar is architectural constraints. OpenAI argued that limiting the solution space can raise agent productivity, because fewer wrong paths need to be explored. The team enforced these boundaries with custom linters, structural tests, LLM-based reviewers, and pre-commit hooks. The third pillar is entropy management, aimed at cleaning up the drift that accumulates in AI-generated code over time, such as naming inconsistencies, stale documentation, and dead code. Cleanup agents were used to scan repositories on a schedule and open PRs to fix problems automatically.
Agents handling the full software loop
In the workflow described by OpenAI, an agent can scan the repository, reproduce a bug, record a failing demo video, implement a fix, run tests, inspect UI rendering, submit a PR, and respond to review feedback in repeated cycles. If the CI pipeline fails, the agent can diagnose and repair the issue on its own. Human escalation happens only when product direction or design judgment is required.
OpenAI also described a “minimize blocking” merge philosophy. In a system with strong testing, monitoring, and quick feedback loops, the cost of waiting can exceed the cost of fixing. The article also cited a LangChain result on Terminal Bench 2.0, where a coding agent improved from 52.8% to 66.5% without changing the model, only the harness. The changes included self-verification loops, context engineering, loop detection, and a “reasoning sandwich.” Together, those examples point to the same shift: engineers are moving away from writing code line by line and toward designing the environment in which agents can work safely and repeatedly.