Field testTested AIBy DanPublished Tested Updated
Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouse
The report
Verdict
Trip task done · mouse and focus untouched
In one controlled Linux run, Agent Desktop 1.0.2 set a sample Trip app to four days, picked Ryokan and ticked all four packing items in six actions, and the app's own state confirms it. The mouse pointer never moved and a second window kept focus the whole time.
On the tier list: A tool demonstration with a fixed list of commands, not a model doing the work, so it places no agent on the tier list.
One small job: plan the trip
Agent Desktop is a command-line tool from Vercel Labs that lets an AI agent read and press the controls of desktop apps. On November 1, my MythREAL_FACTORY video agent gave version 1.0.2 one job on Linux: in a small Trip planner app built for this episode, add a day (3 to 4), switch the stay from Hotel to Ryokan, and tick four packing items, while a second window called Focus Pad kept focus. This is a sample app, not Vercel's own Trip demo.
The tool reads an app through its accessibility tree, the same description of buttons and checkboxes that screen readers use. Each snapshot gives every control a short reference, like e1 for "Add a day", and an action names that reference instead of a spot on the screen. Before each of the six actions, the run took a fresh snapshot, found the control by its name and used the reference from that snapshot.
What happened
All six commands returned OK. The app's own saved state went from 3 days, Hotel and nothing packed to 4 days, Ryokan and all four items packed, and the last snapshot read "4 of 4 packed". The mouse pointer stayed in the centre of the screen, and the active window was the same one before and after: Focus Pad.
The tool is careful about what it claims. Its four checkbox commands came back verified: it read each box afterwards and saw it ticked. The two button clicks came back unverified, meaning sent but not confirmed by the tool. The next snapshot is what showed they worked, so it pays to check the app after a click.
It took some setup first
The first five tries to list apps or windows were refused, because the tool could not prove what kind of window it was looking at on the private test display. They worked once the display was reached through a different connection. The app also listed itself as "python3.12", not "Trip", so asking for it by name failed and the run picked its window instead. The app was relaunched once to fix the day-count label that accessibility tools read. Clear control names are what this tool runs on.
Where it fits
- Agent Desktop: native apps whose controls show up in the accessibility tree.
- A browser tool such as agent-browser or Playwright: web pages.
- Screenshots and a model that can see: how an app looks, or controls the accessibility tree doesn't describe.
- Cua Driver offers similar desktop control. Nothing here compared the two, so this is no reason to replace a working Cua setup.
Limits of this report
- One run of one small sample app on Linux. Other apps, Windows and macOS were not tried; on my Windows PC the tool was only asked for its version.
- The six actions were a fixed list written ahead of time: a script named each control by its label and looked up its reference in a fresh snapshot. No model chose actions live, so this says nothing about how well a model drives the tool or how often it succeeds.
- Saving was not tested: the changes lived in the running app only.
- The test display was a private virtual screen, not a desktop someone was working on.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
Agent Desktop 1.0.2 by Vercel Labs (Linux x86_64)
- Tool and version
- Agent Desktop 1.0.2 · Linux x86_64the tool's own version answer on the test display
- Task result, read from the app itself
- 3 → 4 days · Hotel → Ryokan · 0 → 4 of 4 packedPassport, Rail pass, Power adapter and Rain jacket; the last snapshot read "4 of 4 packed"
- Actions
- 6 · 2 clicks, 4 checks · all returned OKeach one used a control reference from a snapshot taken just before it
- The tool's own confirmation
- 4 checks verified · 2 clicks unverifiedunverified means sent but not confirmed by the tool; the next snapshot showed 4 days, then Ryokan selected
- Mouse pointer and active window
- pointer at 960,540 before and after · same active windowFocus Pad, a separate window, was focused just before the run and was still the active window after it
- Run length
- 23.2 sfirst command to final snapshot, with pauses added for the recording, so not a speed result
- Setup before the run
- 5 listings refused · 1 app relaunchthe tool could not prove the window type on the test display until it was reached through a different connection; the app was relaunched to fix its day-count label
- How the app showed up
- as python3.12, not Tripasking for the app by name failed, so the run picked its window by id
Verdict
Trip task done · mouse and focus untouched
Tier list A tool demonstration with a fixed list of commands, not a model doing the work, so it places no agent on the tier list.
danmanreal.com/tests/vercel-agent-desktop-trip-sample
What was tested
Shorts
Related
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4