01
Human interface
What a person should see and understand.
FROM THE BIC TOOLKITSmall useful things built in response to a slightly broken internet.
BIC UTILITY / FIELD CHECK 003
The Agent Experience Checker
Can an AI system understand the task well enough to act safely?
EXPERIMENT / WEBSITE AFTER THE WEBSITE
One page. Three jobs.
For people responsible for websites: turn a page and a user task into a plan for humans, machine-readable facts and agent actions.
Fictional learning example—not findings about a real shop. Nothing is fetched, sent or saved.
02
Machine-readable knowledge
Facts a system should retrieve without guessing.
Add evidence or an objection
03
Agent-action layer
What an agent needs to act reliably, if action is appropriate.
Add evidence or an objection
YOUR IMPLEMENTATION BRIEF / UPDATES AS YOU ANSWER
Keep the useful page. Fix the missing layer.
This plan reflects your answers. “Not checked” means verify first, not a confirmed defect.
Human interface
- Keep
Keep a plain-language purpose and next step on the visible page.
- Improve
Show choices, costs, dates and consequences before a person commits.
- Keep
Keep the human help and alternative route easy to find.
Machine-readable knowledge
- Improve
Give key facts a canonical source, responsible owner and review date.
- Improve
Expose stable identifiers, units and definitions; do not rely on page layout to convey meaning.
- Improve
Structure conditions and exceptions alongside each fact, including who or what they apply to.
Agent-action layer
- Improve
Define the allowed action, required inputs, validation and expected result before exposing it.
- Improve
Define identity, permissions and a review-and-confirm boundary before consequential changes.
- Verify
Verify first: Are failures, retries and receipts defined? Then: Define failure states, duplicate-safe retry rules, escalation and a completion receipt.
Read or manually copy the plan
Test one change, repeat the same task, and record where a person or agent still gets stuck. This is an experiment, not a claim that websites are disappearing.
WHY THIS EXISTS
A readable help page is not the same as an executable journey.
A person can bridge missing context, hunt for exceptions and ask for help.
An AI system needs the route, rules, branches, consequences and permission boundaries to be explicit.
This check reveals where a task still depends on human guesswork.
AGENT EXPERIENCE UNITSTATUS: READY
PUBLIC EVIDENCE. EXPLICIT RULES. NO GUESSWORK.
Could an agent complete this task?
Answer from the information a system could actually reach. Give the site credit only when the route or rule is explicit.
Nothing you answer here is sent anywhere. The check runs entirely in your browser.
WHAT IT CHECKS
Five conditions for a task an agent can understand.
Findability
Can the official route be found without private context?
Rule Clarity
Are steps, deadlines and consequences stated rather than implied?
Context Handling
Are plan, role, provider and channel branches visible?
Safety & Permission
Are authentication, review and explicit confirmation built in?
Recovery
Can failure, exceptions and wrong-account errors be recovered?
THE BIGGER POINT
The hidden decision tree is the product.
This is a directional self-assessment, not a certification. The useful output is the weak condition and the next fix: the rule, branch or safety boundary that needs to become explicit.