Empathy-first design breaks when the human leaves the room

The reassuring story about AI agents goes like this: they’re just a new kind of user. Your product used to serve people. Now it also serves the software those people send in to act for them. Same good design habits apply, the argument runs. You’ve simply got more visitors.

That story has a serious backer. The Nielsen Norman Group, the UX research firm most product teams cite by reflex, has spent over a year making the case. They describe agent-driven software as “intent-based” interaction, where you state a goal and the system works out the steps, and they call it the first genuinely new interface paradigm in sixty years [verified]. Their prescription is calm: your existing principles hold, the audience just expanded to include agents. Add four requirements for the larger audience. Show what the agent is doing, explain why it chose its path, let a human override at any point, and recover cleanly when something breaks [verified]. This isn’t a niche debate. Gartner projects that 40% of enterprise apps will have task-specific agents built in by the end of 2026, up from under 5% in 2025 [reported].

Then Marty Cagan stood up at the Lenny & Friends Summit on September 11 and named a case the calm version doesn’t cover. Cagan runs Silicon Valley Product Group and is the most-read writer in product management. When he plants a flag, teams reorganize around it. His claim: on scheduling, triage, and coordination workflows where an agent is the primary user, empathy-first research becomes “partially suspect.” Empathy-first research is the discipline of studying real people’s feelings and frustrations and designing for their emotional experience. On agent-run workflows, Cagan says, designing for human emotional resonance while the agent optimizes for task completion creates “competing design objectives” [reported].

Your UX principles carry over when agents join your human users. On the workflows where agents replace them, your most human principle points the wrong way.

Watch the difference between joining and replacing. Take a consumer app with a support flow. A person uses it, and sometimes the agent they’ve delegated to walks the same flow. Empathy for that person is empathy for a user, and the agent riding along changes nothing. NN/G is exactly right here. Now take a triage bot that sorts incoming tickets, or a scheduling agent that books meetings from end to end. On those surfaces the human isn’t a smaller second audience. The human isn’t the party the product serves at all. The agent does the whole job. Design that flow for human emotional resonance and you’re tuning it for someone who never arrives, while your actual user, the agent, just wanted the tickets sorted faster.

So the audience didn’t grow. It shrank past the human. A principle written to serve a person, applied where no person is the user, does more than lose a little force. It aims your effort at the wrong target.

There’s a fair objection, and the NN/G camp already left room for it. They named “dual-audience design” as their open frontier: the case where human needs and agent needs pull against each other and you balance the two. Design the form field for the agent, the navigation for the human. On that reading Cagan hasn’t broken anything. He’s filled in the conflict case the field said was coming, and the position works as intended.

That reading holds only if the conflict is a tradeoff. A tradeoff means both principles still apply and you allocate between them. Cagan is describing something else. He isn’t saying empathy-first matters a bit less on agent workflows. He’s saying it’s “suspect” and “requires inversion.” Underweighting a principle and reversing it are different problems. You balance a tradeoff. You withdraw an inversion. “Principles apply equally to agents” survives the first and contradicts the second.

Here’s the tell. The NN/G position wrote down, in advance, the one finding that would change its mind: a documented case where making software friendlier to agents measurably degraded it for humans. Read that closely. It watches for agents hurting humans. Cagan’s case runs the opposite way, where human-friendly design degrades the agent’s effectiveness. The field strung a tripwire across one side of the door, and the evidence came in through the other. The wording isn’t the accident. Even while planning for conflict, the field still pictured the human as the party at risk. It never pictured a surface where the human isn’t in the audience to protect.

The position has a subtler defense, and it fails on a different axis. Over the summer it already retreated once, from “principles transfer” to “categories transfer.” A Beijing user and a San Francisco user need the same four things, the argument goes, built with different meaning. Reasonable. But that defends a category that needs a different skin. Cagan’s case is a category that shouldn’t be on the surface at all. Empathy-first doesn’t need new semantics for an agent-only workflow. It needs to be gone, replaced by modeling the agent’s task. “Categories hold, implementations vary” has nothing to say about a category that drops out the moment the human leaves.

Hold all of this at its real weight. Cagan’s claim is one keynote, reached through secondhand write-ups, with no transcript, no study, no second source behind it. And inversion is his house move. He finds it everywhere, in competencies that turn into liabilities and roles that flip under AI. A thinker who sees inversions in every domain will see one here whether or not it’s the deepest reading. So this doesn’t refute the NN/G call; it trips the alarm the call installed on itself, from the field’s most-tracked voice, and the call hasn’t answered.

The question worth chasing is narrow and it’s answerable. Go find one agent-only workflow, a triage queue or a scheduling agent with no person in the loop, where empathy-first research measurably improved what the agent did. If that case exists, the calm version wins outright. Expansion logic covers replacement too, and the worry was noise. If nobody can produce it, then “your UX principles apply to agents” was always scoped to rooms that still hold a human. Whether one of your users is still a person is the variable that decides which way the principle points, and the comfortable version never thought to name it.