Type: url
Original: https://marmelab.com/blog/2026/09/24/the-state-of-ai-harness-engineering-2026.html
Status: generated
Language: en
Summary
This Marmelab blog post surveys the state of harness engineering for coding agents in 2026, based on an audit of 246 open-source repositories, a counter-test of 145 large non-AI projects, a shortlist of 11 harnesses that test themselves, and 57 publications. It reports that most projects write agent instructions but few enforce them through hooks, deny rules, tests or evals, and recommends testing and evaluating the harness, keeping it minimal, backing rules with scripts, letting the agent run the app, measuring harness usage and pruning stale controls, while listing approaches the literature found not to work.