Harness Engineering: Making Coding Agents Dependable
For engineers who already use Claude Code, Cursor or Codex and have watched an agent go off the rails on a real repository. You will learn why agents fail on long tasks and build Kite, a small Python harness with permissions, a completion gate, a progress file and an event log, then run it on a real Flask app.
Course Content
Why capable coding agents still fail on real repositories, and which layer — model, harness or repo — owns each failure.
Instructions, scope rules, hard limits and completion gates that hold a single agent session to what you actually asked for.
How to carry work across many short agent sessions, using files and git the harness controls, instead of one long conversation that decays.
How to see what an agent run actually did, diagnose a failed run in minutes, and test harness changes without spending tokens.
Six milestones that build Kite — loop, tools, permissions, completion gate, progress file and event log — and end with a full run on Ledgerly.