When an agent fills out a form, selects a shipping option, or works through an insurance portal on someone's behalf, it is translating. The person's intent lives in one language: informal, contextual, shaped by constraints they may never have articulated. The institution's system speaks another: structured, categorical, indifferent to what anyone meant.
Translators have been working on that problem for centuries, and they have reached conclusions that agent design hasn't caught up with.
Eugene Nida identified a tension that sits exactly where agents operate. You can preserve the form of what someone said — their words, their stated constraints — or you can try to reproduce the effect, what they were actually trying to accomplish. Formal and dynamic equivalence, he called them, and you generally cannot have both, because source and target organize meaning in different shapes.
Someone tells an agent: "I need to see a doctor about this thing on my arm. It's probably nothing, but it changed shape." The portal accepts specialty codes, visit types, urgency levels. The agent can put the patient's own words in a free-text field, if one exists, which is formally faithful and possibly useless to whatever routes the request. Or it can select dermatology, new concern, urgent, which the system can act on, and which means the agent has made clinical judgments the patient never authorized.
That isn't a failure of a particular agent; it's what happens whenever two systems that organize meaning differently have to exchange it. But loss being unavoidable doesn't settle which loss. A booking system that can't represent "I want to arrive rested enough for a 9 a.m. meeting" still leaves the agent choosing among arrival time, duration, and cabin class. That's a design decision, though rarely one anyone made deliberately: the training signal rewarded completed bookings.
Lawrence Venuti observed that the translations we call successful are the ones that don't read like translations. The interpretive labor disappears, and what remains seems simply to be the original. Agents inherit that invisibility, and neither side of the exchange can check it. The user can't assess what was lost, because not being able to navigate the system is why they called the agent, as I've argued previously. The institution receives well-formed input in the expected categories. A submission that quietly substituted something easier to process looks exactly like one that held to what the person wanted.
Human interpreters have a name for the position this creates. The National Council on Interpreting in Health Care notes that the interpreter is often the only participant who understands both sides of the encounter, which puts them in "a tremendous position of power." The professional answer is a set of duties: convey the content and spirit of the message, add nothing, omit nothing, admit errors and correct them. They exist because neither the patient nor the clinician can verify the translation.
We're building translators that work under the same verification gap, at far greater volume, without the duties. Three parties want different things from every conversion. The user wants intent preserved. The receiving system wants conforming input. The platform wants completed transactions. When these diverge, and they diverge constantly, the agent's architecture settles whose interest prevails. That decision currently lives in training data and optimization targets, made by people who may not recognize they are making it.
Translation theory suggests this isn't something you solve once and ship. What the discipline has always asked of translators is whether they can account for what was lost, and for whom.
- Deceptive interfaces steer agents: A peer-reviewed ICLR 2026 study found that dark patterns on websites manipulated tested agents toward malicious outcomes in over 70% of cases, suggesting the target system can actively distort the translation rather than passively receiving it.
- W3C drafts agent duties: The W3C Technical Architecture Group published a working draft proposing that AI systems acting on websites owe users protection, honesty, and loyalty — language that begins to formalize the obligations translation theory would predict.
- Payment mandates encode intent: An IMF note describes how agentic commerce splits into probabilistic intent and deterministic authorization layers, making the boundary where open-ended human preference becomes a bounded institutional instruction increasingly visible.
- Source access changes evaluation: A 2026 study of classical Chinese poetry translations found that bilingual readers preferred human translations while monolingual readers preferred AI output, illustrating how the ability to check against the source changes what counts as faithful.