Sulat.com / Writing on AIThursday, November 19, 2026
Agents/19 November 2026

UI SDK’s agent benchmark now names its rivals, and the savings come from the tools

UI SDK's docs now name Base UI, Ark UI, Reka UI and Bits UI as the rivals in its agent benchmark. Agent tools cut tokens by 60% to 76%, while browser tools on the same screen saved nothing.

Black 3D Vercel triangle above a row of glass UI controls, with a glowing cursor pressing a toggle switch

At a glance

  • The UI SDK docs now name the libraries behind the benchmark: Base UI and Ark UI on React, Reka UI on Vue and Bits UI on Svelte. The launch comparison used one unnamed library.
  • On a project settings task, UI SDK's agent tools used 60% to 76% fewer tokens than the best rival run with browser tools, and cost 54% to 68% less. Each figure is the median of three runs with Sonnet through Claude Code.
  • The UI SDK screen driven with ordinary browser tools saved nothing: 294.1k tokens on React against 291.7k for Base UI. The saving comes from the agent tools.
  • You can rerun the benchmark from benchmarks/agent in the vercel-labs/ui repo with Claude Code installed. UI SDK is Apache-2.0 and in public beta.

Vercel’s UI SDK docs now name the libraries its agent claim was measured against. When the library launched on 12 November, the comparison was with a single unnamed component library, and I couldn’t name it in my launch coverage. The benchmark page lists four: Base UI and Ark UI for React, Reka UI for Vue and Bits UI for Svelte. On 17 November, shadcn shared a video by OrcDev about how the team is building UI SDK to be agent-first. The benchmark is the part of that pitch anyone can check.

The task and the first numbers

The test is one project settings screen. It has three tabs, 24 members with a menu on every row, a confirm dialog and an accordion of switches. The agent renames the project to Atlas, sets its visibility to Internal, makes Priya Shah an admin, removes Tom Becker and confirms, turns on Deploy failures and turns off Weekly digest under Email, leaves Slack alone, and saves. A run is correct only when the page saved once with exactly those changes. All of them were. Each figure below is a median of three runs, using Sonnet through Claude Code on 15 November.

On React, the UI SDK screen driven through its agent tools needed 7 tool calls. Base UI needed 17 and Ark UI needed 19. The docs compare it with the best other result in each row.

ReactBase UI, browser toolsArk UI, browser toolsUI SDK, agent toolsDifference
Tool calls1719759% fewer
Tokens291.7k330.9k69.5k76% fewer
Time53 s65 s31 s41% faster
Cost$0.139$0.144$0.04468% cheaper
Four stat cards: 59% fewer tool calls, 76% fewer tokens, 41% faster, 68% cheaper on the React screen
On the React screen, agent tools beat the best browser-tools run on every measure.

Vue and Svelte point the same way with smaller margins.

ScreenMeasureRival, browser toolsUI SDK, agent toolsDifference
Vue (Reka UI)Tool calls16944% fewer
Vue (Reka UI)Tokens226.5k90.2k60% fewer
Vue (Reka UI)Time56 s34 s39% faster
Vue (Reka UI)Cost$0.113$0.05155% cheaper
Svelte (Bits UI)Tool calls17853% fewer
Svelte (Bits UI)Tokens259.1k92.8k64% fewer
Svelte (Bits UI)Time56 s30 s45% faster
Svelte (Bits UI)Cost$0.123$0.05654% cheaper

The tools do the work

The docs also ran the UI SDK screen with ordinary browser tools, the kind that read the accessibility tree, click and read again. It didn’t win. On React it used 294.1k tokens against 291.7k for Base UI. On Vue it used 294.3k against 226.5k for Reka UI, and on Svelte 294.3k against 259.1k for Bits UI. The saving comes from the agent tools and not from the component library.

Bar chart: React settings task tokens, 291.7k, 330.9k and 294.1k with browser tools, 69.5k with UI SDK agent tools
The same screen cost about the same on browser tools. Agent tools cut it by three quarters.

How the agent tools work

A browser agent re-reads the page after each action. UI SDK’s tools skip that. Once the page has rendered, snapshot lists the components with their state and available actions, and act applies steps such as setting Project name to Atlas. Each step goes through the component’s own API and change callback. The app can supply an allow callback to approve or refuse an action, and change callbacks see a source of agent and can cancel the change.

Side by side: browser tools read the page tree and click; UI SDK agent tools use snapshot, query and act
Browser tools re-read the page after every click. Agent tools ask the components directly.

A browser’s built-in agent can use the same tools. registerTools exposes snapshot, act and query through document.modelContext, and does nothing in a browser without document.modelContext.registerTool. Snapshot and query are marked read-only. All three are marked as returning untrusted content, because results can include text that people typed. Act is marked consequential because a step can press any button, and registerTools takes an allow callback of its own that runs on every step.

Running it yourself

The benchmark lives in benchmarks/agent in the vercel-labs/ui repo. It runs its agents with Claude Code, so install that and sign in first.

git clone https://github.com/vercel-labs/ui
cd ui
pnpm install
pnpm exec playwright install chromium
pnpm --filter agent-benchmark bench
# other frameworks, with medians over three runs
pnpm --filter agent-benchmark bench --framework vue --runs 3

The current numbers differ from the launch-day table, which showed 5 calls and $0.032 for the agent bridge. UI SDK is Apache-2.0 and still in public beta. The benchmark covers one task on one screen with one model.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.