Samyak Jain
Design

The reader that never skims

Your product has a second user now. It reads every word, pays for each one, takes instructions literally, and acts on someone else's behalf. Designing for it is a different craft.

Samyak JainOctober 26, 20268 min read

In 1994 a software engineer named Martijn Koster proposed a small text file called robots.txt. Website owners could use it to tell automated crawlers which pages to leave alone. It was an early admission that websites had gained a second kind of reader, one that was not a person.

Within a few years that reader mattered as much as the first. Search engines decided who got visitors, and a whole discipline grew up around writing pages that a crawler would understand. The crawler changed how the web was built, and it only ever read.

The second reader arriving now does more than read. It fills in forms, calls APIs, compares prices and spends money, all on behalf of a person who may never open your product. I think most product teams are about where web teams were in 1995: aware that this reader exists, and not yet designing for it.

The interface matters as much as the model

The clearest evidence comes from a 2024 paper by researchers at Princeton. They built a system called SWE-agent to fix real bugs in open-source software, and they ran one comparison that every product manager should know about.

They took a single model and gave it the same bugs twice. The first time it worked through an ordinary command line, the same one a human engineer would use. The second time it worked through an interface designed around how the model reads and acts: a file viewer that showed a hundred lines at a time, a search command that returned a short summary, and an edit command that checked the code for errors and refused to save a broken change.

Same model, different interface

Share of software issues resolved on SWE-bench Lite, GPT-4 Turbo (Yang et al., 2024)

Same model, plain command line11%Same model, interface designed for it18%
Figure 1. Nothing about the model changed between the two runs. The authors call the designed layer an agent-computer interface, by analogy with the human-computer interface.

The model resolved 11% of issues with the plain command line and 18% with the designed interface. That is a 64% improvement from interface design alone. The command line was built for people, who glance at long output and pick out what matters. The model read all of it, lost its place, and made edits it could not see the effect of.

How this reader differs

Good interface design starts with knowing the user. This user is unusual.

Two users of the same product

A personAn agent
ReadingSkims, guided by layout and colourReads every word it is given, and pays for each one
MemoryRemembers last week's visitStarts fresh unless someone hands it notes
InstructionsFills gaps with common senseTakes the description literally
When stuckGets annoyed and leavesKeeps trying, sometimes in circles
ChoosingSwayed by brand, habit and designSwayed by the description and by what worked last time
AuthorityActs for themselvesActs for someone else, within limits it may not fully know
Figure 2. Most interface conventions exist to help a reader who skims. They do little for one who reads everything.

Two of these differences drive most of the design consequences.

The first is that attention has a price. A person skims a page for free. An agent reads every token, and each one costs money and uses up some of its limited working memory. A response that returns two thousand words where two hundred would do makes every later step slower, more expensive and more likely to go wrong.

The second is that the description is the product. A person chooses a tool because of its brand, its design, or a friend’s recommendation. An agent chooses from a list of tool names and one-paragraph descriptions. Whatever those paragraphs say is, to the agent, what your product is.

What good design looks like

Anthropic’s engineers published a detailed account of what they learned writing tools for Claude. Their advice lines up closely with the differences above.

Offer whole tasks. A typical API exposes small operations: list the users, list the events, create an event. A person writing code strings these together. An agent has to work out the sequence afresh each time, and each extra call is another chance to slip. A single tool that finds a free slot and books the meeting does better than three that must be called in order.

Return meaning. Their team found that returning names and plain descriptions in place of long internal ID codes reduced the model’s rate of made-up answers. The model reasons in language, and an ID gives it nothing to reason with.

Keep answers short by default. In one of their examples, a concise version of a tool’s response used 72 tokens where the detailed version used 206, with the agent able to ask for more detail when it needed it.

Write errors as directions. “Error 400” stops an agent, or sends it in circles. “The date must be in the future; today is 13 November” lets it fix the problem on the next try. For this user, the error message is the onboarding.

Treat descriptions as your most important copy. The same team reports that small, careful changes to tool descriptions produced large gains in how often the model completed its task. In most companies this text is written once by an engineer in a hurry. It deserves the attention a landing page gets.

Someone else’s agent

There is a further difference, and it concerns authority more than reading.

The agent in your product is acting for someone. Economists have studied this arrangement since the 1970s under the name of the principal-agent problem: one party acts for another, with different information and sometimes different interests. The classic examples are managers acting for shareholders and lawyers acting for clients. An AI agent booking a trip for a traveller fits the same pattern.

Three parties, two relationships

The personSets the goalJudges the resultTheir agentReads everythingPays per tokenYour productTools and docsPermissionsGoal, limits,budgetTool callsApprovals,receiptsResults, errors,prices
Figure 3. Your product sees only the agent. It still has to respect limits that the person set.

This puts a new kind of requirement on a product. It has to know what this agent is allowed to do, and behave sensibly at the edges. A few questions are worth settling early:

  • Which actions can the agent take alone, and which need the person to confirm? A useful line is whether the action can be undone. Searching and drafting are safe. Paying and deleting are not.
  • Can the person set a limit, such as a budget or a scope, that your product enforces even if the agent forgets?
  • When something needs approval, does the request reach the person in a form they can judge in a few seconds?
  • Is there a record afterwards of what the agent did and why?

Standards for all of this are forming. The Model Context Protocol gives agents a common way to find and call tools, and payment companies have published protocols for purchases made by agents. The details will change. The need for a product to know whose agent it is talking to will stay.

A usability test for the second user

Product teams test with people as a matter of course. The equivalent for this user is easy to run and rarely done.

Take a fresh model with no knowledge of your product. Give it only what a real agent would get: your tool descriptions and your documentation. Give it ten tasks that your customers actually want done. Then read the transcripts.

Where the agent hesitates, calls the wrong tool, or misreads a response, you have found a design problem, in the same way you would if a person got lost on a screen. Measure three things: how often the task is completed, how many tokens it took, and how many times the agent had to retry. Those are the conversion rate, the time on task and the error rate of the second user.

The first user has not gone away. A person still decides whether the outcome was good, and still chooses which agent to trust. But more and more, the path to that person runs through a reader that will never see your homepage, and reads every word of everything else.