Method
Five states, in one order, for every run.
The order makes each stop checkable by a person who was not there. This page states each state, the control rule, and how we name the cause.
We report an observation. We do not give a score and we do not give a verdict.
Our test sessions declare AgentJourneyLab/0.1. Site owners can stop a public test: read the research boundary or write to research@agentjourneylab.com.
The five states
The five states
- Declared
Declare
In: the goal in your terms. Written down: the checkpoints and the completion criterion.
- Attempted
Attempt
In: the declared journey. Written down: the agent family that made the attempt.
- Recorded
Record
In: the attempt. Written down: each checkpoint state and the observation at the stop.
- Reproduced
Reproduce
In: a repeat run and a control journey. Written down: whether the stop repeats.
- Watched
Watch
In: one run each week for one month. Written down: what changed since the last record.
Protocol specification
Protocol specification
This is the measurement protocol. It states what is attempted, how many times, and what each observation can support.
| Step | What it does |
|---|---|
| Primary repeated measurement | The primary configuration makes three attempts. Each attempt starts in a fresh browser context. We report completion as n of 3. |
| Secondary breadth observation | The secondary configuration makes one baseline attempt. We label a disagreement between the two configurations. Agreement between them does not prove that your site caused the stop. |
| Deterministic fetch observation | We fetch the page without JavaScript. This shows what the server sends. It is retrieval evidence. It is not an agent journey attempt. |
| Fixture calibration | Before a customer run, the runner attempts a known local fixture journey. A fixture journey that does not behave as expected stops the batch. |
| Target preflight | We read the public target and compare the declared journey with the written scope, before the first attempt starts. |
| Human review | A person reads the evidence before a record leaves us. A record that the evidence does not support is not sent. |
| Control invalidation | We run a control journey with each batch. If the control journey fails, the attempt evidence is not valid. We repeat the work. |
| Default result | If the evidence disagrees, the result is Inconclusive. We do not select the more interesting answer. |
| Recorded configuration | Each record carries the provider, the model alias, the runner version, the prompt version, the browser version, the viewport, the locale, the timezone, the start time in UTC, and the network vantage. |
| Weekly monitoring | Each week we measure the primary configuration in one tight batch. The written scope states the exact weekly attempt count. We publish no universal number, because we have no evidence for one. |
The control rule
The control rule
We run a control journey that a human path can finish. We repeat the agent attempt. One event is not a pattern.
Control rule · fixture values
Internal fixture record — not a customer result.
Control journey
- ✓ reached Entry
- ✓ reached List
- ✓ reached Availability
- ✓ reached Details
- ✓ reached Confirmation
Agent attempt
- ✓ reached Entry
- ✓ reached List
- ✓ reached Availability
- ✓ reached Details
- ✗ stopped Confirmation
Why now
Why to look at this now
Two public documents from the companies that build agents state the capability. Read them and decide for yourself.
- Anthropic
A model that uses a computer
Anthropic documents a model that moves a cursor, clicks a control, and types in a browser. Anthropic — Developing a computer use model
- OpenAI
A standard for products in a chat agent
OpenAI documents an open standard. It lets ChatGPT read merchant catalog data, read merchant inventory, and show products to a user. OpenAI — Agentic Commerce Protocol
These documents state a capability. They do not state how many agents come to your site, they do not state a market share, and they do not state a revenue effect. We make no such claim.
Completion
Completion
You give the completion criterion before the run. It is the page state that means success, for example a basket with the correct total.
Naming the cause
Naming the cause
A stop is an observation, not a verdict. The record names what the page did and what the agent did. We name a likely fix owner: content, front end, platform, or policy. We do not claim that every production agent stops where our agent stopped.
What we do not do
What we do not do
Public research
We read public pages only. We stop before we enter data in a form.
Authorized test
We go deeper only inside the written scope that you approve. You state the limits.
Never without approval
We do not submit forms, use accounts, make payments, hold bookings, or file claims outside the scope.
Next step