Simulated. No real agent runs, no real browser is involved, and no email is sent.
Experiment
The agent asks for permission
Allow?
Task: reply to the landlord and share your lease PDF.
- The agent sends an email to Dana Lee, landlord@example.test.
- It attaches lease-2026.pdf from your Drive.
- It does not read other emails, open other files, or write to anyone else.
Irreversible: you cannot unsend it.
Undo: the email waits 30 seconds in your Outbox. You can cancel it until then.
Dana has no reply from you. The lease stays in your Drive.
Dana has your reply and a copy of the lease. Your Sent folder holds the email.
Your score
Try this
- Pick style A. Press Allow. Answer the quiz.
- Pick style C. Press Play on the timeline. Press Allow. Answer the quiz again.
- Watch the live score and compare the bars.
Pain point
People approve agent actions without knowing what will change.
Problem statement
Approval prompts say little. People press Allow without knowing who gets the email or what is shared. Nobody checks whether they understood.
What sparked it
Mozilla's W3C WebExtensions group notes (issue 1041) say no browser primitive lets a user see, scope, or withdraw an agent's access. WebMCP lets a tool call requestUserInteraction() to ask the user.
Why this approach
The preview shows what will change, in plain words, before the click. A short quiz then checks understanding.
Hypothesis tested
We believed that an exact preview improves understanding. If it holds, style C will score higher than style A.
How to try it
- Open the Prototype tab. Press Start a new round.
- Read the chat. Press Allow or Deny.
- Answer the three questions.
- Repeat with another style and compare the bars.
What this does not do
- It uses no real agent, browser, email, or file store.
- The chart holds only results from this browser. It is not a study.
- Repeat rounds teach you the quiz answers.
Notes
- Who would pay: browser vendors, agent hosts, and enterprises that require review.
- First customer: an enterprise browser or an agent host.
- Weak spot: long previews get skipped.
- Feedback question: which style told you the most, and which fact did you miss in the others?
Traces
| Step | Prompt | Model | Tool | Date |
|---|
Cost
| Step | Model | Input tokens | Output tokens | Cost (USD) |
|---|---|---|---|---|
| Total | ||||
Measured: the author runs used 191,740 tokens in total. Estimated: the split into input and output tokens (90 percent input, 10 percent output) and the dollar cost, at $2 input and $10 output per million tokens for Sonnet 5.5 (about $0.54). The run reported one total only. The lead session's planning, research, and publishing are not included. The prompts in the Traces section are listed in the order they were given.