Ten harness patterns that you can apply to a real repo

Reliable coding-agent work depends less on perfect prompts than on engineering the system around the model. A harness supplies context, tools, checks, and feedback so agents can work within a codebase’s constraints; people set goals and guardrails, while agents carry out tasks. Drawing on experiments with Codex and Claude Code, the guidance aims to reduce time spent reconstructing project context and correcting avoidable errors. It presents ten practices spanning repository documentation, just-in-time context, durable tools, discoverable validation, bounded feedback loops, selective visual checks, intent-focused prompts, proportionate planning, turning repeated mistakes into guardrails, and keeping workflows stable as models change.

The first practice is to make the repository an agent-readable knowledge base. A concise root-level AGENTS.md should explain the project, architecture, constraints, conventions, entry points, and areas that require human approval, then link to deeper documentation rather than trying to contain everything. The code shows what a system does, but commit messages and regression-test docstrings can preserve why decisions were made and which failures a safeguard prevents. Agents can draft those explanations, with humans checking their accuracy and feeding recurring corrections back into the guidance. This turns documentation, history, tests, and validation tools into durable context: each task can begin with relevant knowledge, and mistakes can improve the system instead of recurring.