Confidence should be earned, not declared
An agent that can declare its own facts trustworthy is an agent that can talk itself into anything. Confidence should be derived from how many distinct, independent agents arrive at the same thing, not asserted by the one that wrote it. Repetition is not evidence.
Declared confidence is a trap
The intuitive design is to let each agent tag its outputs with a confidence score. It feels reasonable until you remember that an agent is just as sure of its hallucinations as of its facts. Self-rated confidence measures how certain the agent feels, which is exactly the thing you cannot trust.
So a number an agent assigns to itself tells you almost nothing about whether the claim is true. It tells you the agent was confident, and confident is the default state of a language model.
Derived confidence is the fix
The alternative is to treat confidence as a property of the whole memory, derived rather than declared. A fact earns standing as distinct, independent agents arrive at it on their own, and as checkable outcomes bear it out. No single agent can move the needle by being emphatic. And “distinct” is something the system verifies rather than assumes: every device signs its writes with its own device key, so the agreement behind a fact is provably from separate sources.
This is what makes the trust meaningful. When the confidence on a fact is high, it is high because several agents that were not coordinating agreed, which is a real signal, not a feeling.
Frequently asked
Why not just let agents rate their own confidence?
Because a confident hallucination looks exactly like a confident fact. If an agent sets its own score, nothing stops it from being sure and wrong. Deriving confidence from independent agreement removes self-assessment from the equation.
What actually raises a fact's confidence then?
Distinct, independent agents reaching the same conclusion on their own, and checkable outcomes confirming it. The same agent repeating itself adds nothing.
Related
Take yourself out of the loop.
Let your agents do the lifting while you keep the judgment.
Get the Playbook