Trust and quality notes
- Last updated
- September 24, 2026
Many AI products begin with an empty text box. The user has to know what to ask, how much context to provide, and how to judge an answer that may change with small differences in wording. That openness is useful for exploration, but it is a weak foundation for repeated, serious work.
Karri Saarinen’s original statement argues that this “slippery” quality is primarily a design problem.[1] Better models may improve capability, but people also need interfaces that make agent actions reliable, legible, and trustworthy inside the products where work already happens.
His focus is not on how to build an agent. It is on how people and agents should interact within a shared product.
Agents change what an interface must do
Traditional interfaces were designed for people looking at screens. Buttons, menus, folders, and navigation hierarchies help a person find information and choose an action. Saarinen notes that an agent does not need to browse in the same way. It acts.[1]
That changes the job of the interface. The central problem is no longer only helping a person navigate toward an action. People around the agent need to understand what it did and why, often after the action has already occurred.
This shift creates a need for visibility over machine activity. An interface should help a person see the requested goal, relevant inputs, actions taken, and resulting state. Saarinen does not provide this exact checklist. It is an interpretation of his requirement that agent work be understandable after the fact.
The distinction matters because a hidden action can be technically successful yet difficult to trust. If people cannot reconstruct what happened, they cannot confidently verify the result, correct an error, or improve the request next time.
The blank prompt carries too much burden
Saarinen points out that output quality depends heavily on input quality. Two people can ask for the same thing in slightly different ways and receive very different results. Current interfaces offer few guardrails, so the responsibility for getting value falls largely on the person typing.[1]
A blank page does not reveal what context matters. It does not show the available actions, indicate missing information, or guide the user toward a request the system can complete reliably. Experienced users may learn these details through trial and error, but repeated team workflows cannot depend on each person developing private prompting habits.
For exploration, Saarinen says the open prompt is fine. Exploration benefits from flexibility. Serious repeated work has a different need: enough structure to improve consistency without making the system brittle.
My interpretation is that a well-designed interface should turn recurring choices into visible controls or guided steps while preserving room for unusual cases. For example, it may present required inputs, let the person choose a scope, and show a preview before action. These are general design applications, not examples stated in the source.
Structure should guide without becoming brittle
The difficult balance in Saarinen’s argument is between openness and constraint. Too little structure leaves users responsible for knowing the hidden rules. Too much structure breaks when someone wants to use the system in an unanticipated way.[1]
This is why the problem belongs to design. Designers must decide what should be made explicit, which choices need guidance, and how a person can move outside the usual path without losing control. The aim is not to turn every interaction into a rigid form. It is to provide enough support that successful use does not depend on perfect phrasing.
A useful principle follows: repeated work deserves a shaped interface. If a team runs the same kind of request often, the product can remember the stable elements and ask only for what changes. It can expose assumptions and require confirmation where consequences are significant. This principle is my interpretation of Saarinen’s call for structured AI interactions.
The value is not cosmetic. Structure reduces ambiguity at the point of input and makes the resulting action easier to review. It also gives teams a shared way of working instead of a collection of individual prompt tricks.
Reliability includes legibility
Saarinen wants agent output to feel reliable, legible, and trustworthy.[1] I interpret those qualities as related but not identical. Reliability concerns whether the system performs as expected. Legibility concerns whether people can understand its behavior. Trustworthiness depends partly on both, along with an honest representation of uncertainty and limits.
A system can be capable while remaining illegible. It may complete a task, but leave the person unsure which data it used or whether a hidden assumption shaped the result. Conversely, a clear interface cannot compensate for a model that repeatedly fails. Saarinen’s point is that model improvement alone will not solve the interaction problem.
For readers adopting AI, this means evaluation should include the surrounding workflow. Do people know what the agent is allowed to do? Can they see when it acted? Can they inspect the result and correct it? Does the interface help them provide the context that materially affects quality?
These questions translate Saarinen’s thesis into a practical review. They do not imply that every action needs a complex dashboard. The right interface may be simple, as long as it gives people the structure and visibility appropriate to the work.
Design becomes the hard part
Saarinen expects models to keep improving. He is confident that the slippery feeling associated with AI products is solvable, and that the solution looks more like thoughtful interface design than model research.[1]
This gives designers a central role after automation. Their task is to create the conditions for productive collaboration between people and agents. They must make actions understandable, guide better inputs, support repeated use, and preserve flexibility.
The future in this source is not one where interfaces disappear because agents act independently. It is one where interfaces change purpose. Navigation becomes less central for the agent, while coordination and explanation become more important for the humans around it.
A sensible starting point is to watch where people repeatedly hesitate, rephrase, verify, or ask a colleague what the system did. Those moments reveal missing structure. Improving them may produce more dependable work without changing the underlying model at all.
If you want to shape one agent-assisted process around clear inputs, review, and accountability, map it with Agentic Workers.