The Agent Harness, Not the Model, Decides Whether AI Workflows Hold Up
lxfater · x · 2026-11-21
The author argues that unreliable AI work is usually blamed on model intelligence, but a new Google course and a Stanford assignment both point to the outer "harness" — the braking and fallback layer — as the real determinant of reliability.
- Google's example: two agents both book a restaurant; one checks the calendar first, the other guesses and gets lucky. Same score today, but the guesser fails on the next run.
- Stanford's open assignment bans frameworks and requires writing a harness from scratch; of 88 automated test points, 48 don't call the model at all, instead feeding empty replies, malformed output and errors to test fault tolerance, plus a budget cap that hard-stops runaway runs.
- The author, who has run dozens of Skills in workflows, says hard-coded budgets and check scripts beat repeated prompting.
Bottom line: the model is the engine; the chassis and brakes decide whether it can go on the road. Links to the Google Cloud course and the GitHub assignment repo are included.
More from coding & agent
- Grok Bot template turns one idea into a coded animated video with music — nima_owji · 2026-11-21
- OpenResearch Intelligence Index: first open benchmark for end-to-end autoresearch — BenBlaiszik · 2026-11-21
- Fake font files in 21 GitHub repos hide malware auto-run via VS Code tasks — JeremyCMorgan · 2026-11-21
- Hugging Face ships harness adapters: RL-tuned Qwen3.8-27B gains on Terminal-Bench 2.1 — yb2698 · 2026-11-21
- Japanese tech lead argues code design matters more, not less, in the AI coding era — 4310sy · 2026-11-21
- Google open-sources stitch-skills: 16 design skills for AI coding tools, 11k GitHub stars — lxfater · 2026-11-21