Teaching AI to See

Giving any agent eyes

Your tests pass, but the UI is broken

ui-vision gives an AI agent eyes so it can catch what assertions can't — the button that shifted 40px, the text clipping behind a card.

First, proof this thing is real.

Assertions check data and state
They don't catch the button that shifted 40px left.
The screen goes unseen
Text clipping behind a card sails past a green run.

ui-vision is real, public, and AI-built

An itself-AI-generated Amplifier bundle, committed overnight on Feb 13, 2026 — commit 2092877, co-authored by Amplifier, verified via git log.

So what exactly do assertions miss?

12
files in the initial commit
728
lines inserted in one commit
00:17
Pacific, Fri Feb 13, 2026
1
commit — AI-generated by Amplifier
The Blind Spot

Assertions check data and state — they never see the screen

The README says it plainly: your tests pass but the UI is broken. Visual breakage sails past a green test run because the model was never looking.

ui-vision closes that gap by letting the model look.

The model literally sees the screenshot

ui-vision runs Playwright MCP in vision mode. Each screenshot returns as a base64 image the model can see and describe — not just a file path.

Seeing is only half of it — scope is the other half.

# behaviors/ui-vision.yaml args: - "@playwright/mcp@latest" - "--caps=vision" - "--headless" - "--image-responses" - "allow" # browser_navigate -> go to URL # browser_take_screenshot -> model SEES # browser_snapshot -> accessibility tree

Three bounded specialists, one question each

Instead of one open-ended browser agent, ui-vision ships three specialists — each asking a single, bounded question with a matching recipe.

And the whole set drops into any bundle.

visual-auditor
Does this look good?
regression-checker
Did my changes break anything?
accessibility-scanner
Can everyone use this?

The whole capability is a single 31-line YAML

Vision server, three specialists, three recipes — the entire capability composes into any bundle through one 31-line behavior YAML.

Which means eyes stop being a bespoke integration.

31
lines in behaviors/ui-vision.yaml
3
bounded specialist agents
3
ready-to-run recipes
1
behavior YAML to compose it all
The Payoff

Any agent can now look at what it built and judge it

Eyes are no longer a bespoke integration — they're one line of YAML away. Drop the behavior into any Amplifier bundle and the agent inherits sight.

That changes what verification depends on.

The Takeaway

Verification stops depending on a human staring at the screen

When agents can see their own work, composable sight closes the gap between "tests pass" and "it actually looks right."

Sources

Sources & Research Methodology

Feature status: Shipped & Public

Data as of: Feb 2026 — bundle cpark4x/amplifier-bundle-ui-vision (4 commits, Feb 13 → Mar 15, 2026).

Every claim independently re-derived from live git / gh / grep / wc / read — nothing trusted from the input file:

Primary contributor: Chris Park (cpark4x) — all bundle commits; Amplifier bot co-authored the initial commit.

Gaps: A wider "autonomous build then crash" framing exists in derived decks but could NOT be traced to any surviving source repo, commit, or PR; those figures are excluded here. This deck presents only the verified ui-vision bundle.

More Amplifier Stories