deepsense.ai won an OpenAI Codex hackathon, then turned the lessons into ten harness patterns for real repos
Ten harness patterns that you can apply to a real repo
After winning an OpenAI Codex hackathon with HyperCode, a repository intelligence layer, the deepsense.ai team tested harness patterns in their own setups with Codex and Claude Code. The playbook covers three layers: making the repo an agent-readable knowledge base via a short AGENTS.md and tests that explain why, compiling context just in time for each task, and closing validation loops so recurring agent mistakes become durable guardrails in docs, tests, or linters.
Your coding assistant can read your code. What it cannot infer on its own is why the code is structured the way it is, what tradeoffs were made, which abstractions are intentional, and which parts are legacy landmines.