swarmwallet

Inspiration

Comparison shopping in 2026 still means having to open a dozen tabs, refreshing pages, and hoping you didn't miss anything, so we figured we could build something that helps makes this process exponentially quicker through browser agents, while saving our wallets with full transparency.

However, deploying a lot of browser agents are slow and so optimizing that process is what inspired us to use Wafer.

What It Does

You type a product. In seconds, a swarm of browser agents fans out across Amazon, Best Buy, Target, Walmart, eBay, and Etsy simultaneously each one running a real browser, extracting real prices. While the swarm is running, you watch it happen: a live swarm map, streaming prices as they come in, and Browserbase replay iframes so you can see exactly what each agent saw. No black boxes.

Once results are in, Wafer reads across all the pages and produces a final verdict — best deal, price spread, any outliers. Snowflake handles semantic routing, stores 90-day price history, flags suspicious outliers, and powers a /memory dashboard that gets smarter every swarm.

There's also a /benchmark route that runs a side-by-side synthesis comparison against Gemini, purely for measuring speed, quality, and cheap tradeoffs. This is just to show benchmark comparison.

How We Built It

The architecture is a Next.js frontend talking to a Node WebSocket server, which orchestrates the whole swarm. The flow looks like this: our swarm plan kicks off parallel agents (one per retailer), each powered by Stagehand for navigation and Wafer for extraction. Results then feed into Snowflake for a semantic intelligence layer, and then Wafer synthesizes everything into the final verdict.

Here are our key pieces:

Browserbase: remote Chrome sessions with captcha solving and proxy rotation Stagehand: handles navigation and extraction, with a custom Wafer client (disableAPI: true) so it routes through our LLM instead of the default Wafer (through wafer pass - Qwen3): every agent extract and the final synthesis Snowflake + Cortex: persistent swarm history, vector embeddings, warehouse-level insights

Challenges

In the very beginning, we took way too long to think about our product amassing over 6 hours until finally deciding on swarm.wallet. Definitely took a lot our time away from our health.

Browser VMs have real cold-start latency which is about 10 to 30 seconds. So we prewarmed sessions and tuned concurrency caps to make it feel fast enough to actually us, but this burned through our browserbase usage when testing.

Anti-bot measures, per-site DOM structure differences, and captchas all needed custom handling per retailer.

Limit of 3 concurrent requests was also a bottle neck for us considering we were conducting a swarm of agents.

Prompt to path generation was fairly tricky in the sense of we originally gave our orchestrator to generate a task plan for which sites to choose and search, which could have made the plan a little too complex and too long to launch so many different browsers. Therefore, we really had to think about the best optimization for most popular sites.

Accomplishments

We're proud that this does real parallel browser shopping and its not just like a simulation. We think the live agent generation is super cool.

Wafer runs on every extract and every synthesis step, and it's fast enough to work at swarm scale. That was the core bet of the project, and it paid off especially because its such a cheap model.

The Snowflake memory layer means each swarm makes the next one smarter which meant that we get better routing, better context for the verdict, richer history for the user. This was something super cool especially because we were thinking about how we could incorporate each memory layer to improve the agentic behavior.

And the Wafer vs. Gemini benchmark is actually useful data, not just a demo flourish.

What We Learned

Browsers are cool and the integration layer when APIs don't exist. Parallelize at the browser level, minimize LLM hops, and use a fast model that won't bottleneck your swarm. Wafer made that viable.

Snowflake is pretty cool.

What's Next

We could definitely test this with other things beside just products. Buying food or grocery delivery could be a great idea for whats next (need to be wary of web scraping permissions though).

On the inference note, we can definitely improve our speeds and times if we had a higher rate limit.