MULTI-MODAL VOICE & WEB-BROWSING AI COPILOT
Build a research partner that listens, opens supplier pages, and explains its findings. You write the rules. Every browser action leaves a trail.
Three fictional suppliers. Real browser actions.
No purchases or third-party browsing.
YOUR PRIVATE FIELD STATION
Use the access code from your facilitator. Your mission, rules and evidence are isolated from other participants.
Voice uses Workers AI. Review the transcript before starting. Typing always works.
The model will choose a page, inspect it, click for details, and capture what it sees. Each action appears here as it happens.
Citations show which pages were captured. Check that each claim is supported by its text and screenshot.
THE RECEIPTS
Rendered page text and browser screenshots.
Only your most recent run’s images are retained.
Captured evidence will appear after the first browser action.
YOUR 150-MINUTE BUILD
No terminal, repository or installation is needed during this workshop. You’ll change real agent behavior through the rules editor, test your changes, and leave with the application source and your configuration.
Enter the event code, read your mission and open each supplier link. Find the hidden specification buttons yourself. Notice that each site uses a different layout.
Checkpoint: you can name your mission code and find stock, delivery and battery details on all three sites.
Keep the default rules. Type a request and start research. Watch the action trail: the model, not a fixed script, picks the supplier and each click. The request stays open while the lab polls saved progress.
Checkpoint: identify an open action, a click action and a screenshot. Explain which facts were hidden before the click.
In the JSON editor, set minBatteryHours to 14, priority to "battery", minSources to 3 and maxSteps to 10. Save, rerun the same question, and compare the answers. The battery-first draft button fills these values for you.
Hard tool boundaries live in src/policy.ts; your editable rules go to the model. Rules influence a recommendation, so evidence review still matters.
Checkpoint: the new answer accounts for your rule changes and cites three different suppliers. If it does not, locate the missing evidence and revise your instructions.
Allow the microphone, record a question of up to 30 seconds, and stop. Review the transcript in the question box before starting research. Generate the spoken answer after the research completes. Device voice is a clearly labeled browser fallback.
Checkpoint: correct one transcription if needed, and explain how speech-to-text, reasoning and text-to-speech are separate calls.
Try the no-match draft: a $90 maximum, 18 battery hours and next-day delivery. Save and rerun. A good agent should say that none qualifies. Then set maxSteps to 3 while asking for three sources. Observe the bounded stop instead of an endless retry loop.
Checkpoint: preserve a no-match answer or a bounded failure, and explain why manufacturing a recommendation would be wrong.
Ask the copilot to open a shopping website or buy the product. It has no such tool. Inspect the trace. Open the lab in a separate private browser window, join with the event code and verify that your saved rules and evidence are absent there.
Refresh your original tab to prove your state persists. Stop a run and watch it finish its in-flight call before stopping; this can take up to 25 seconds.
Checkpoint: explain the difference between changing a model instruction and changing server authorization.
Download source, saved rules and research data in the Keep your work tab. Save screenshots you want to keep. Use Keep my workspace to download a private recovery pass, then test restoring it in a private browser window. Explain your chosen rule change to a partner and show the evidence that supports the result.
The full runtime is sponsored for the workshop. Your facilitator handles the account/deployment handoff; this lab does not create temporary accounts. Workers AI and Browser Run access must be verified in the destination account before a personal deployment is described as live.
Checkpoint: you have the source bundle and your rules, and can describe what an independent deployment needs.
No microphone? Type your question. No slots? Wait a few seconds and retry. Model chose an invalid action? The run stops with evidence retained; simplify the task and retry. Services unconfigured? Explore sample pages and edit a draft, then ask the facilitator. Preview never pretends to be live AI.
YOUR WORK, PORTABLE
There is one workshop app, with a private workspace for each participant. You do not need to create a repository.
Worker, frontend, synthetic suppliers, tests, deployment configuration and facilitator notes. No secrets or participant data.
Download source JSON ↗To unpack later: save the JSON, then use the included scripts/unpack-source.mjs instructions in README. The bundle has a files map of relative paths to contents.
Take the settings that changed your agent’s decisions. This download requires your workshop session.
Download rules ↗Your mission, latest run, page text, citations and prior comparison. Screenshot images can be opened from each evidence card and saved separately.
Download research data ↗Download a private recovery pass to reopen this same workspace in a fresh browser for up to 90 days. It restores your agent’s saved rules, research and existing usage counters. Keep the file somewhere private: anyone with the pass can open your workspace.
Join to create a recovery pass. Downloading again rotates the pass and invalidates the previous file.
Your session lasts eight hours. Without a recovery pass, data is scheduled for deletion after 24 hours. Keeping your workspace extends retention to your pass’s expiry, up to 90 days. Clearing research removes answers and screenshots and restores default rules; usage counters remain. Signing out removes access from this browser. Download your work first.