trajectoryrl caught a subtle regression in my agent's system prompt that only showed up on multi-turn conversations. single-turn evals would've missed it entirely. the multi-turn testing is the differentiator, that's where real agents actually break
multi-turn is where agents actually break and single-turn evals give you false confidence. the state accumulation across turns is exactly what a real conversation stresses
the multi-turn regression that single-turn misses is the exact class of bug that ships to production and then blows up on a real user's third message. anyone shipping conversational agents without multi-turn eval is flying blind and doesn't know it.
