Complete ARIA relationships
What it checks
Composite widgets (menus, listboxes, tabs, grids) require specific child/parent roles and attributes. Missing pieces break the structure agents traverse.
Why it matters
Agents choose which field to fill by matching the user’s intent against the accessible name of a control in the accessibility tree, then write to that control by reference. An input whose name comes from a properly associated <label for>, aria-label or aria-labelledby appears as, for example, textbox “Email address”. An input with only a visual placeholder, or an adjacent unassociated div, appears unnamed. Natural-language lookup then fails, and the agent either skips the field or writes the right value into the wrong control.
Evidence
Form labels and ARIA naming for agent form-filling
The full chain is first-party documented. Anthropic’s find will ‘Search for elements matching a natural-language description such as “search field” or “add to cart button”’, and returns refs in the same tagged format as read_page. form_input then ‘Set a form element’s value directly’ by that ref. The accessible name is literally the lookup key [anthropic-browser-use-tool]. Playwright MCP’s snapshot contents are documented to include ‘form elements (textboxes, checkboxes with accessible names)’. browser_fill_form and browser_type act on snapshot refs [playwright-mcp-snapshots, playwright-mcp-repo]. accname, a W3C Recommendation, fixes the precedence: aria-labelledby over aria-label over the native <label>, alt or title [w3c-accname].
HTML-AAM confirms that ‘label: provides accessible naming for form controls’ [w3c-html-aam]. browser-use keys on the textbox, combobox, checkbox and searchbox roles, plus the a11y properties ‘focusable, editable, settable’ [browser-use-clickable-elements]. Behavioural corroboration: across 300+ trials, agents engaged only where ‘semantic button overlays or off-screen text labels are present’ and ignored purely visual affordances [machine-readable-ads-paper].
Limits
Pixel-based agents (OpenAI computer use, Gemini Computer Use, Anthropic desktop computer use) fill forms from visual layout and do not require programmatic labels at all [openai-computer-use-guide, gemini-computer-use-docs, anthropic-computer-use-tool]. Adoption data shows agents already cope with widespread absence. Only about 35% of mobile inputs get their accessible name from a <label> [web-almanac-2025-accessibility], yet agentic form-filling demonstrably works across the live web. Placeholder text, nearby text and CSS position all provide fallback signal, and Anthropic’s find is explicitly a fuzzy natural-language matcher rather than an exact accessible-name lookup.
No public benchmark isolates label presence against form-fill success rate, so the effect size is undocumented even though the mechanism is airtight. Score presence of an accessible name from any accname-recognised source, not the literal <label for> element.
How it scores
The whole chain is first-party documented. Anthropic’s find locates “elements matching a natural-language description such as ‘search field’ or ‘add to cart button’” and returns refs. form_input then sets a value by that ref. The accessible name is literally the lookup key, so a control named only by a placeholder is not addressable by intent. A named vendor, a named tool and a stated mechanism is grade A. It scopes to programmatic agents: pixel-based ones (OpenAI computer use, Gemini Computer Use, Anthropic desktop computer use) fill forms from visual layout and need no programmatic label at all.
Sources
- Browser use tool (browser_toolset_20260801) — Anthropic, vendor-doc (verified 2026-08-21)
- Snapshots — Playwright MCP — Microsoft, vendor-doc (verified 2026-08-20)
- microsoft/playwright-mcp README — Microsoft, vendor-doc (verified 2026-08-21)
- Accessible Name and Description Computation 1.1 — W3C, spec (verified 2026-08-20)
- HTML Accessibility API Mappings 1.0 — W3C, draft-spec (verified 2026-08-21)
- browser-use ClickableElementDetector source — browser-use, vendor-doc (verified 2026-08-20)
- Machine-Readable Ads: Accessibility and Trust Patterns for AI Web Agents interacting with Online Advertisements — arXiv (Nitu, Mühle, Stöckl, 2025), study (verified 2026-08-20)
- Web Almanac 2025 — Accessibility chapter — HTTP Archive, dataset (verified 2026-08-20)
- Computer use — OpenAI API guide — OpenAI, vendor-doc (verified 2026-08-21)
- Computer use — Gemini API — Google, vendor-doc (verified 2026-08-20)
- Computer use tool — security and prompt injection guidance — Anthropic, vendor-doc (verified 2026-08-20)