Shyam Sarma / Wayfinder
ContactWayfinder
Watch an AI agent research real products and build the comparison.
Wayfinder gets one practical job: compare three familiar laptops for a travelling AI engineer. It visits the official Apple, Microsoft, and Samsung pages in a real browser, records only what it can verify, moves into LibreOffice Calc, and leaves a workbook you can download.
Wayfinder / isolated computer
The Apple product page will open here.
You can watch, but the agent controls this computer.Technical traceVisible computer actions from the live run
This is the lower-level execution log. It is separated from the explanation above so the demo remains readable for people who do not need implementation detail.
- 00No run yet
The technical trace appears after Wayfinder starts.
How Wayfinder is built.
A live agent loop connects the page, the model, and a temporary computer. The finished file is checked before anyone can download it.
- 01Next.js + Vercel
The page starts a server-side mission and receives live progress events.
- 02Agents SDK + Luna
The model decides which visible computer action to take next.
- 03E2B desktop
A temporary browser and spreadsheet app execute those actions.
- 04Verify + deliver
The workbook is checked, then offered for download if valid.
Architecture, tools, and tradeoffs.
What each piece does, and why it was included in this demo.
Agent runtimeOpenAI Agents SDK with GPT-5.6 Luna.
The agent loop runs on the server. Luna can use a computer tool for visible actions and a separate progress tool for short, viewer-friendly updates. The progress feed is a deliberate report from the agent, not its private reasoning.
Computer and streamE2B supplies a fresh desktop for each mission.
Wayfinder acts through a temporary E2B desktop with a browser and LibreOffice Calc. The page displays that desktop and receives curated progress events while the server coordinates the run. The desktop is closed when the mission ends.
Sources and trustOfficial pages are evidence, not commands.
The comparison is limited to three selected product pages. Their contents are treated as product evidence rather than instructions for the agent. When a specification is not clearly published, Wayfinder marks it as missing instead of guessing.
Verification and storageA workbook has to pass checks before download.
The server inspects the saved workbook for the expected product rows, source links, and recommendation. Only a valid result is stored for download. Upstash Redis supports run admission and records, while Vercel Blob holds the finished file.