Agent Lenses

Status: FUTURE. Today agents see HTTP JSON and paste text.

The sensory gap

Traditional software agents primarily see:

• APIs
• databases
• files
• text
• DOM structures
• logs

Users see pixels. Many production bugs are pixel bugs: overlap, off-canvas, zero contrast, modal behind header, chart clipped, button present in DOM and display:none in the visual stack.

Loop

Agent: Think -> Act -> Observe API -> Observe database -> Observe visual result -> Evaluate -> Correct -> Repeat

EE VIZN is the extra observe step. Pastes stores the judgment so the next agent does not repeat the same miss.

Worked example

1. Agent PATCHes a dashboard widget. API returns HTTP 200.
2. DOM contains button.save.
3. Computed style and screenshot show the button covered by a sticky filter bar.
4. Text-only agent marks the task done.
5. Lensed agent writes a Pastes run: screenshot asset, defect tag occlusion, evaluation fail, links the prompt and model.
6. Correction: CSS order or breakpoint. Re-test. Promote only if visual eval and API eval both pass.

That distinction is strategically important: technical success is not user-visible success.

Implementation notes (when built)

• Playwright (or equivalent) as the browser observer.
• Never screenshot pages that display live secrets; use fixtures.
• Store only signed object URLs in Pastes.
• Visual fail does not auto-revert payments; it opens a Command task.