1 comment

[ 0.26 ms ] story [ 17.7 ms ] thread
Hi HN,

Tried running some multi-step agent workflows recently, and honestly — things broke more than I expected.

Not because the model was “bad”, but because:

small errors compound across steps outputs drift slightly each time one wrong assumption early= everything downstream breaks What surprised me is this: model intelligence didn’t matter as much as consistency

Even when using mm2.7, which works fine in isolation, reliability across chained steps becomes the real bottleneck So now I’m wondering:

What models are people actually using for agent workflows? Are you optimizing for intelligence, speed, or consistency? How do you deal with failures across steps? So now I’m wondering:

What models are people actually using for agent workflows? Are you optimizing for intelligence, speed, or consistency? How do you deal with failures across steps?