When agents are unsure, they should tell you, not guess
A language model has no built-in tell for the difference between something it knows and something it invented, so left alone it states both in the same confident voice. An agent worth trusting flags its uncertainty out loud — marking a shaky claim as shaky and asking instead of guessing — so you can see the gap before you act on it.
A guess and a fact sound identical
The dangerous thing about a language model is not that it is sometimes wrong. It is that it is wrong in the same fluent, certain voice it uses when it is right. There is no waver, no hedge, no tell. When it does not know something, the path of least resistance is to produce a plausible answer and state it exactly as confidently as a verified one.
That means the burden of detecting uncertainty cannot fall on the reader noticing a change in tone, because the tone never changes. If an agent is going to be unsure honestly, the uncertainty has to be a property the system tracks and reports, not a feeling the prose is supposed to convey.
Surfacing the gap instead of filling it
The honest move is to leave the gap visible. When an agent does not have support for a claim, it should say so — mark the claim as thin, or ask rather than invent — instead of reaching for the most likely-sounding completion and presenting it as settled.
HiveMind builds this into how memory works. A fact’s standing is derived from how many independent agents arrive at it and whether checkable outcomes bear it out, not from the confidence of whoever wrote it down. A claim that only one agent asserted, with nothing else behind it, does not get to wear the same certainty as one that several uncoordinated agents confirmed. The thinness is recorded, not hidden, so when you look you can see which facts are load-bearing and which are a single agent’s guess.
That distinction is what lets you step back safely. You are not asking your agents to be right about everything — that was never on offer. You are asking them to be honest about what they do not yet know, so that the doubt reaches you while you can still do something about it. An agent that guesses silently takes a decision away from you. An agent that surfaces its uncertainty hands it back.
Frequently asked
Why don't agents just say when they're unsure on their own?
Because fluent prose carries no signal about certainty. A guess and a fact come out in the same confident voice, and the model has no internal meter that separates them. Surfacing uncertainty has to be designed into how the system stores and reports what it knows, not left to the agent's tone.
Isn't an agent that keeps flagging uncertainty just annoying?
Only if it flags everything. The point is to mark the claims that are actually thin — the ones one agent asserted with nothing backing it — and let the well-supported ones stand quietly. Calibrated doubt is useful; constant hedging is noise.
Related
Take yourself out of the loop.
Let your agents do the lifting while you keep the judgment.
Get the Playbook