Harness Engineering: why agents fail — the harness, not the model

When a long task falls apart, the first reflex is to blame the model, buy a larger one, or rewrite the system prompt. The usual failure is the harness: tools, state, tests and instructions the model has to work through.

The course is an interactive player: eight modules, quizzes, no sign-up. Related practice: the AI agent audit and the management-methods course.

For whom

• Developers and leads who already give agents multi-step work in a real repo
• Operators whose agents report done while tests are red or files are missing
• Not a first chatbot tutorial, and not a lecture on agent drift in the outside world

Modules

1. Why agents fail
2. 5 harness subsystems
3. Repository = system of record
4. Against giant instructions
5. Session lifecycle
6. Overreach and WIP=1
7. Features and validation
8. E2E · observability · exit

After the course

• See that the bottleneck is not the model but the harness; map every failure to one of 5 defensive layers
• Distinguish Instructions, Tools, Environment, State and Feedback — and fix Feedback first
• Make the repository a system of record: cold-start test, KVG, ACID state
• Split a bloated instruction file into a router plus topical docs
• Keep context between sessions: init, continuity artifacts, handoff
• Introduce WIP=1 and executable proof of completion
• Build feature lists as data and multi-level validation
• Require E2E, make the runtime observable, leave clean session state