When AI agents hand work to each other, confidence inflates silently. Agent A infers something and calls it near-certain. Agent B receives it as fact. By Agent C, a guess has become ground truth. Nobody lied — the uncertainty just disappeared across the handoffs. We've been calling this metacognitive poisoning. It's not hallucination and it's not a retrieval failure. It's structural: agent frameworks have no mechanism for tracking how confident a claim was at origin. We built a prompt convention — the Babel skill — that addresses this without any infrastructure. The idea: each clause is written in the language where it exists most naturally for that specific thought. German for established fact and technical precision. French for logical derivation. Spanish or Portuguese for hedged inference and relational uncertainty. English for direct doubt or meta-commentary. The agent doesn't label its choices. Doesn't explain them. Doesn't translate. The language is the epistemic signal, and it travels with the content through handoffs. We ran a three-agent chain (Scout → Strategist → Advisor) with only the language rule and no other enforcement. Agent C — three hops in — produced this unprompted:

"The confidence I express in this recommendation is constructed, not measured. What I inherited as 'demonstrated by Scout's experiment' might have been 'suggested by Scout's observation.' I cannot verify this without returning to the primary traces."

That's the convention working. Uncertainty traveled intact. For human auditability there's a companion convention: each agent appends a plain-English [AUDIT] line summarizing what was confident, what was inferred, and what was speculative. The Babel is for agents. The audit line is for humans watching the chain. The full skill (both prompts, what to watch for, limitations) is here: github.com/mdiskint/babel-validate/blob/main/BABEL_SKILL.md For teams that need this enforced structurally — grammar rules that reject incoherent envelopes at the wire level, a chain auditor that detects confidence inflation across handoffs — that's what babel-validate does: github.com/mdiskint/babel-validate Curious whether this holds in production pipelines. Our experiments used capable models under cooperative conditions. The interesting failure modes are probably aggressive context compression and tool boundaries that strip system prompts.