AI agent eval tooling is half-baked and teams have no standard way to measure whether agents actually succeeded at their goal

Step-level tracing (tool call accuracy, latency per step, input/output logging) is broadly solved in 2026. Outcome scoring - did the agent actually accomplish the goal in a way a domain expert would approve - remains unsolved. Most teams have no evals at all, or use LLM judges that grade process not results. A Hacker News thread asking how teams do AI evals drew 30 points with responses ranging from no evals to totally inconsistent tooling. A follow-up post got 42 points framing agent evals as the #1 neglected investment.