1/5 We ran SWE-bench 200,000+ times to get statistical confidence in agentic evals. The main lesson wasn’t about prompts or models.
It was: agentic evaluation is an infrastructure problem.

→ View original post on X — @ai21labs,