agent harness

One night our agent fleet produced at least 9 wrong or almost-wrong claims the harness caught before they shipped — and at least one reached the running server before a second layer of defense masked the impact. An honest account of the harness — probes at the authority source, verifiers separate from generators, hooks that block, typed state over prose — making individually fallible agents collectively trustworthy, with the line drawn exactly where the evidence stops.

A newer, higher-benchmarking model was available at the same price. I rolled my chief-of-staff agent fleet back to the older version anyway. Here is the evidence on both sides, including the part that disagrees with me, and why for agentic systems harness fit beats model tier.

A single-day census of what our production multi-agent harness caught: at least 9 wrong or almost-wrong claims caught before they shipped, and at least one that reached a running server before it was caught. Disclosed counting rule, per-item record including the miss, and an honest statement of what the numbers do and do not support.

We built ASIF for a simple reason — one-shot AI is not enough; real AI work needs a control plane.