Beehive: AI-native Resell Platform

Looking for resellers, buyers, traders? Beehive does it for you.

Comment

Inspiration

I bought a Saturday ticket and a Sunday ticket to a three-day music festival, then decided I wanted a three-day pass instead. To swap, I had to sell both tickets and find someone selling a three-day ticket, all at the same time. Otherwise I could sell my tickets and end up with nothing. It took days of posting, messaging, and spreading the word around. I got lucky and found two buyers and a seller. Later I learned some people had been selling three-day tickets at 50% of the price, but sadly I never saw those listings. Missed opportunity.

College students go through this with concert tickets and sublets all the time, usually through Facebook Marketplace, Instagram notes and word of mouth.

What it does

Beehive helps people find each other - with enhanced speed and scope that you wouldn't get through word-of-mouth or a Facebook Marketplace listing.

Beehive is where you tell an agent what you have and what you want, in plain English. It searches other people’s requests for you, and it asks you questions when something is unclear, like your lowest acceptable price. If nothing fits yet, it waits and checks again when someone posts something new.

The key feature is that it can match more than 2 people. In my example above, I had to look for 2 buyers and 1 seller simultaneously, or the plan wouldn't work. Beehive can handle it, looking for a combination of requests that all fit together to satify everyone.

When the agents have found a working arrangement for all involved parties, you'll be notified and given the contacts details of the other users you're trading with, and finalize the transaction off-platform.

How I built it

Frontend setup using React + Vite + Clerk (Auth) Backend setup using Node.js + Hono Gemini API for agents

The chat with your agent streams live, using server-sent events (SSE). The agents turn free-text requests into structured fields and apply rules to fill gaps (for example, no price listed means the price is negotiable). Agents are told never to offer a number outside what the user said they’d accept.

To check the agents behave, I wrote evaluation tests. They run the agents through scenarios and a separate model judges whether each message broke a rule. Results go to a Sentry dashboard, which is how I found and fixed the cases where the agents misread vague requests.

Challenges we ran into

Arrangements with three or more people were the hardest. Each person has to see only their own part of the deal, and the arrangement can only go ahead if everyone approves their end of the deal.

What we learned

Vague input is the hard part of agents. People leave out the details that matter, like their lowest price or their dates. Getting AI agents to act smartly on this data is very hard.

Observability platforms like Sentry make that a lot easier.

Sentry Challenge

In this project, Sentry let me

  • Watch internal agent-to-agent negotiation in a production setting, crucial for fixing the nasty bug where Agent A claims to be waiting for Agent B's response, while Agent B claims to have not spoken to Agent A at all. Sentry logs gave me the insight that agents were referencing stale versions of each other, causing this to happen.
  • Monitor AI agent evaluation scores, so I know when an agent broke a rule, such as lying to the user.
  • Keep track of LLM costs and failure rates across different tool calls (Agents view).
  • Create custom dashboards that let me observe the step-by-step process behind how an agent finds a working arrangement for its user.

Sentry made me notice the unreliability of my AI Agents' behavior, motivating me to spend more time on perfecting them instead of other tasks that were less important (i.e. UI).

I found Sentry to provide more out-of-the-box tools than Datadog, and have less overhead than Grafana. Really nice.

What's next for Beehive: AI-native Resell Platform

  1. Location-based matching. Users should be able to pick a point on a map and search only for listings a radius around that location.

  2. Trust scores. Flag scammers and view peoples' reputations.

  3. Text messaging through Photon. Your agent gets a phone number you can text, so you don’t need to open the app, making the resale process even more passive/effortless.

BTW, if the app doesn't work, it's probably because I ran out of Gemini AI tokens

Built With

Share this project:

Updates

Submission history