There are six words that strike fear into anyone debugging a user interface: “It looks weird on my laptop.”
For years, the next step was a small interrogation. Which browser? What size is the window? Is that before or after you clicked save? Then screenshots became easy to share, and the interrogation got shorter. Now multimodal models can look at those screenshots too, which makes the image feel almost magical: point an agent at the broken page and ask what is wrong.
This works surprisingly often. It also creates a new temptation—to confuse seeing the evidence with understanding the bug.
The six-word bug report
A screenshot is excellent at preserving visible state. It captures clipped text, a modal hiding behind a header, a mobile layout that never became mobile, or the tiny red badge everybody forgot to mention.
That matters for AI agents. Benchmarks such as VisualWebArena exist because many web tasks are visually grounded. The DOM might say an element is present while the screenshot says it is covered by a cookie banner. Both facts are true. Only one matters to the person trying to click it.
What the image knows
The useful surprise is that a screenshot can be more honest than the application's structured output. A component tree describes what the software intended to render. The pixels show what escaped.
Pixels also preserve relationships that get awkward in logs: this button is too close to that label; this chart legend uses the same color twice; this error message exists, but three screens below the action that caused it.
The missing second
Images are terrible at time. They do not tell you that the button flickered before it disappeared, that the total changed after a retry, or that the user pressed Back twice. They rarely contain intent. A perfectly rendered “Delete complete” screen is still a bug if the person clicked Archive.
They are also selective. The crop may remove the URL, the system clock, another window, or the one browser extension doing something mischievous. An agent that sounds confident while filling those gaps is not debugging. It is writing fan fiction about CSS.
A better evidence packet
The future bug report is probably not a form with twelve mandatory fields. It is a small evidence packet assembled automatically:
- the screenshot and a few seconds of interaction history,
- the user's last meaningful action and expected result,
- viewport, browser, route, and feature-flag state,
- relevant console and network failures,
- a privacy filter before any of it leaves the device.
Give that packet to a multimodal agent and the screenshot becomes genuinely powerful. Not because pictures replaced bug reports, but because pictures finally joined the rest of the evidence.