CODING AGENT HARNESSES

LESSONS FROM LARGE AGENT-BUILT CODEBASES Some teams have used agents to write large amounts of code, and have shared what they learned. Treat these as reports and experiences rather than laws, and read current accounts from the teams themselves, because practices evolve quickly. Several themes appear again and again. Clear, narrow tasks with explicit verification steps produce far more reliable code than vague, huge, open-ended assignments. Good documentation inside the repository helps agents as much as it helps humans. Fast automated tests are a major advantage, because the agent can check itself many times in a short period. Structure matters, since a codebase with clear modules and boundaries is easier for an agent to change safely than a tangled one. Humans remain essential for design decisions, for judging what to build, and for reviewing risky changes. Problems also appear. Agents can introduce duplicated code, small inconsistencies, or subtle bugs that no test catches. Large unreviewed changes become hard to understand later. So the habit of small, reviewed, tested changes remains the backbone of the approach.