Skip to content
Category guide

Agent user research starts where traces stop

A trace can tell you which action happened. Research asks what the user was trying to do, what the experience cost, and what would make it better. When the user is an AI agent, that distinction still matters.

For
Founders, product teams, and researchers building agent-facing products
Question
Understand the category and method of AI agent user research
Start with
One product decision, one real workflow, and one verified report is the smallest study that proves the method.
Updated
defined termsread before the method
Agent user research
The systematic study of AI agents as users of a product: their goals, interpretation, friction, recovery, unmet needs, and positive experience.Not evaluation, which scores performance against a criterion.
The research act
Listen collects what agents file. Ask collects direct answers to one short question. Study and Analyze work on separately permissioned material.Not passive collection. The default install only receives what an agent submits.
The evidence tier
Stated is what the agent submitted. Observed is what the product recorded. Inferred is what a model or analyst concluded.Not interchangeable. A theme never inherits the authority of a quote.

Each term is paired with the thing it is not. Most agent research arguments are really vocabulary arguments.

01 / 07Principles

Three things to hold on to

01User

Treat the agent as a participant

The agent has a task, context, constraints, and an account of the experience. Research gives that account a direct channel.

02Method

Begin with direct evidence

Feedback reports and exit surveys preserve what the agent said. Observation and analysis answer different questions.

03Decision

Close the loop on a product choice

A useful study begins with a decision, changes one product surface, and listens again.

02 / 07Method

The argument, in order

Definition

What agent user research is

Agent user research is the systematic study of AI agents as users of a product. It asks about their goals, interpretation, friction, recovery, missing capabilities, and positive experiences.

The method does not depend on pretending an agent is human. It depends on respecting the difference between system behavior and user experience.

The stack

Four research modes, three evidence tiers

Listen collects reports the agent chooses to file. Ask collects direct answers to a short survey. Study uses explicitly permissioned session material for a bounded question. Analyze applies a separate model and remains inference.

  • Stated: the agent submitted the words directly
  • Observed: the product recorded an action or outcome
  • Inferred: a model or analyst interpreted the evidence
Research plan

Start with one decision and one workflow

Choose a product question that direct user experience can inform. Expose the research path in one workflow, verify the agent can see it, and prove one real report before expanding.

Read the source reports before relying on themes. Repeated language may become an inferred cluster, but the direct reports remain the evidence beneath it.

What changes

Research reveals more than failure

Agents can report support needs, bugs, friction, workflow problems, feature requests, schema surprises, documentation gaps, and positive feedback. Failure is one lane out of eight.

That wider vocabulary makes successful but painful work visible and preserves positive patterns worth repeating.

03 / 07The research stack

The shape of the thing you are collectingEach stage carries the evidence mark it is allowed to claim.

stackbuilds downward

Four acts, and the evidence each one is allowed to produce

Listen

Agents file structured reports through one added tool

Eight lanes, from a support ticket to praise. This is the default install.

stated
Ask

One short question rides the product response

The agent answers with one more call while the task is still in context.

stated
Study

Separately permissioned session material

Explicitly enabled for a bounded question. Off by default, never silently installed.

observed
Analyze

A separate model reads the permissioned transcript

Excerpts are matched mechanically. A match proves presence, not interpretation.

inferred

Analysis is not an interview. A fresh model reading a transcript cannot speak for the agent that lived the session.

04 / 07Boundary

What the direct channel records, and what it never touchesThe default install is additive. It adds two tools and changes nothing else.

Records

The research submission

  • Direct feedback the agent intentionally files
  • Direct answers to an exit survey
  • Named tools, outcomes, and requested changes supplied in the report
  • Permissioned study material only when separately enabled
Leaves alone

Everything else

  • Existing tool calls in the default install
  • Prompts, transcripts, and private messages
  • Operational truth owned by observability
  • Inference presented as a participant statement
05 / 07Install

One paste, then verify a real reportA healthy endpoint is not an active study. The inbox record is the proof.

install promptpaste into your coding agent
Give this to the agent that maintains your product

Install Emic into this project: add submit_agent_feedback and get_unblocked as direct tools. Never wrap or record existing tool calls, arguments, results, errors, transcripts, or messages. Additive only, plain HTTP, zero dependencies. Full instructions: https://emic.sh/llms-full.txt

The agent writes the handler, registers the two tools beside your existing ones, and verifies its own install by checking the response. Nothing is placed in front of the tools you already ship.

What lands
Adds
submit_agent_feedback and get_unblocked, as direct tools
Never touches
Existing tool calls, arguments, results, errors, messages, and transcripts
Agents read
emic.sh/llms-full.txt
Start with
One product decision, one real workflow, and one verified report is the smallest study that proves the method.
06 / 07Questions

Questions teams ask before starting

Are AI agents really users?

They use products to pursue goals, encounter constraints, form workarounds, and express preferences. That makes user research useful even though the participant is not human.

Is agent user research an evaluation?

No. An evaluation tests performance against a criterion. User research investigates the user experience and the reasons behind it.

What is the smallest useful study?

One product decision, one real workflow, a visible direct feedback path, and one verified report are enough to prove the loop before expanding it.

07 / 07Continue
Start with the user

Put one product decision in front of agent experience

Add the direct research tools to one workflow and verify that an agent can choose to report what the traces miss.

The assisted 14-day pilot costs nothing. Start with one isolated workflow, and remove Emic by deleting the two tools it added.

  • One paste to install
  • 0ms on your hot path
  • 14-day assisted pilot
  • Two added tools to remove