One Year of AI Agent Development: 440,000 Lines of Code and the Shift of the Developer's Role to Control and Improvement

• As developers delegate coding to AI agents, their core work shifts from coding to agent control and improvement.
• A closed-loop structure was built where design, rules, and reviews are performed in an isolated environment outside the repository, feeding results back into design and rules.
• Rule audits analyze repeatedly violated, unused, and non-violated rules to propose new or abolished rules, currently managing 274 standard rules and 299 patterns.
• Code quality is verified by tracking post-merge code survival rates and complexity trends, applying the principle of not trusting agent self-reports based on a case where the resolution rate dropped from 87.1% to 73.0% in the SWE-bench Pro sealed environment.

Redefining Harness Engineering

• This is an approach that redesigns the entire development process to match the surge in code productivity, rather than just agent configuration techniques.
• Similar to OpenAI's 'Humans steer. Agents execute.' principle, it focuses on restructuring the development process for defect-free structural growth.
• It is a system designed to address the increased deployment instability upon AI adoption highlighted in the DORA report.

Verification and Observation System

• Mechanical verifications such as type checking, linting, and static analysis serve as the basic safety net, while a separate layer where LLMs judge meaning and context is established.
• During PR reviews, an issue ledger is used to maintain context across iterations, and convergence is ensured by pointing out only serious issues in unchanged code.
• Code survival rates are directly measured, with 93% of additional PR lines surviving one week after merge and 92% surviving on day 24.

The Developer's New Role

• Instead of writing code, time is invested in defining specifications and resolving trial-and-error issues, with 3 to 4 months spent establishing specifications containing approximately 10,000 types of visualizations.
• Patterns, issues, and techniques extracted from design reviews and the review process are accumulated in a catalog to reduce understanding debt.
• The absence of code structures and syntax theories that are easy for machines to read is presented as an open question.