Decision 01
Don’t confuse agent failure with customer friction.
Agent constraint UX finding
An automated browser failing does not prove a human customer experiences the same problem.
UX Scout
Understand the journey beyond your product
Overview
UX Scout is an AI-powered customer journey tool that investigates how goals can be accomplished across the public web. Give Scout a goal, not a predefined path. It independently navigates the web and returns the journey, evidence, constraints, and defensible UX findings.
Role
Product Strategy, UX, and Development
Timeline
48-hour sprint · August 2026
Tools
React/Vite, Node/TypeScript, OpenAI Agents SDK, Playwright MCP, Chromium
01 · The opportunity
The customer journey doesn’t begin at your website.
Product and UX teams often evaluate experiences they own: expected funnels, predefined flows, and individual touchpoints. But customers pursue goals across a much larger ecosystem. What if AI could discover those paths autonomously?
02 · The prototype
Give Scout the goal. Not the route.
-
01
Define
Describe what the customer is trying to accomplish.
-
02
Explore
Scout controls a real browser and independently determines how to pursue the goal.
-
03
Investigate
Scout returns the path, evidence, constraints, and defensible UX findings.
03 · Designing the agent’s judgment
The challenge wasn’t browsing. It was judgment.
Decision 02
Finding nothing is a valid finding.
0 UX findings
Scout isn’t rewarded for manufacturing criticism when the evidence doesn’t support it.
Decision 03
Verify meaning, not just facts.
“Newest”
Newest released?
Newest announced?
Scout checks ambiguous qualifiers before claiming completion. If meaning can’t be verified, it returns a partial result instead.
Decision 04
More autonomy isn’t always better.
Keep exploring? Stop with what’s verified?
Stopping rules and turn limits keep Scout from wandering indefinitely while preserving useful partial results.
Section 04
Real journeys changed the system.
Testing exposed problems I couldn’t anticipate from the initial prototype. Each failure became an input to the next iteration.
05 · Evaluation
Testing the concept against real journeys.
6
Runs
3
Usable
1
Complete
2
Evidence-backed UX findings
| Journey | Result | Usable | Runtime | UX findings |
|---|---|---|---|---|
| Product information | Complete | Yes | 1:47 | 0 |
| Local service comparison | Partial | Yes | 2:18 | 1 |
| Airline policy: open start | Partial | No | 1:50 | 0 |
| Airline policy: supplied start | Partial | No | 2:10 | 0 |
| Entertainment / ticketing | Partial | Yes | 2:39 | 1 |
| Product recommendation | Partial | No | 2:05 | 0 |
- Complete
- Answered the requested goal.
- Partial
- Answered only part of the requested goal.
- Usable
- Despite incompleteness, returned enough verified information to be useful.
06 · How Scout works
Under the hood.
User goal
UX Scout
OpenAI agent
Playwright MCP
Chromium
Public web
Structured evidence
Journey report
Technology
- React + Vite
- Node.js + TypeScript
- OpenAI Agents SDK
- Playwright MCP
- Chromium
07 · Outcome
The idea outgrew the approach.
What I learned
Playwright MCP could investigate individual experiences, but it couldn’t reliably follow a customer journey wherever that journey led.
Scout worked when browser automation could access the experience. But real journeys crossed sites that blocked, challenged, or otherwise prevented automated navigation, leaving gaps Scout couldn’t resolve.
What that means
The opportunity still exists. This just isn’t the right approach to solve it end to end.
The original problem remains: product teams have limited visibility into customer journeys that extend beyond the experiences they own. UX Scout showed me that browser automation alone isn’t enough to document those journeys reliably. Solving that problem would require a different approach, or a combination of approaches.