Anthropic and OpenAI both published posts on agent harness design last month and arrived at the same conclusion. Agents need an external checker that gives concrete feedback. Without it, the agent praises its own mediocre work and moves on. https://goodeyelabs.com/insights/evaluation-is-the-load-bearing-part [Translated from EN to English]

→ View original post on X — @randal_olson, 2026-04-02 13:00 UTC