Friction

Home page
Tasks
Paths
Annotations
Results

Inspiration

Traditional usability testing is valuable, but slow, expensive, and hard to run continuously.

We wanted usability testing to feel more like automated software testing: launch a swarm of AI users, watch them use your product, and see exactly where they struggle.

That became Friction.

Because every click tells a story.

What it does

Friction sends autonomous AI personas through a live website and gives each one a task to complete.

As they navigate, Friction captures:

where they hesitate or fail
screenshots and interaction evidence
repeated issues across personas
severity-ranked UX findings

The result is a replayable UX stress test, with graph and gallery views, that clearly shows where users struggle, why it happened, and what to fix.

Our long-term loop is fully automated:
Deploy → Stress test → Find friction → Suggest solutions → Create pull requests

How we built it

Agents
Multimodal AI agents visually inspect the website, choose actions, interact with the page, and record evidence.

Analysis
Findings across persona runs are grouped to surface repeated friction while filtering out false positives and intentional design patterns.

Frontend
React, TypeScript, Vite, Tailwind CSS, and shadcn/ui.

Realtime
Server-Sent Events stream live agent progress into the control room.

Pipeline
Website → Personas → Agent runs → Evidence → UX findings

Challenges we ran into

Making agents behave like real users instead of exploiting hidden page information
Distinguishing real UX issues from intentional product or business decisions
Running many agents concurrently without excessive memory or API cost
Streaming many live agent sessions reliably through SSE

Accomplishments that we're proud of

Built autonomous agents that visually test real websites
Combined findings across multiple independent personas
Grounded UX findings in screenshots and real interaction traces
Built a live interface for watching agents test a product in real time

What we learned

Giving an AI a browser is easy; making its failures meaningful is much harder
Task completion alone is not enough — hesitation, backtracking, and confusion matter too
Multiple independent personas produce stronger signals than a single AI critic
Agent systems quickly become infrastructure problems involving concurrency, streaming, memory, and cost

What's next for Friction

Scale to 15+ concurrent agents with generated personas and simulated locations
Add stronger cost guardrails and false-positive filtering
Turn persona runs into Playwright E2E tests
Automatically create pull requests from validated findings
Develop complex browser annotations, where arrows, shapes, and other visuals are drawn