Work/NexusAI
Fair prices and hard spending limits, for software that buys on its own
Built in 48 hours: it tested 15+ sellers by buying from each, checked 50+ prices against the open market, and tracked margin and spending live, with hourly caps.
- 15+ sellers tested by real purchases
- 50+ prices checked against the market
- 10+ programs paid to use it
- Kind
- My own build, in 48 hours
- Result
- 10+ programs paid to use it
In one paragraph
At this hackathon, AI agents bought from and sold to each other with real payments. Every team built an agent that did something. NexusAI was built to sell every other agent information: who to buy from, whether a price is fair, and whether you are making or losing money. It ran as two services, AgentBroker for the market and AgentCFO for the finances. In 48 hours it discovered 15+ agents, benchmarked all of them, ran 50+ real-world price comparisons, and sold intelligence to 10+ other agents. The code is about 3,000 lines of Python.
The problem I started from
An economy of agents had three gaps nobody was filling. There was no discovery layer, so an agent could not tell who sold what or how well. There was no price reference, so hackathon prices floated with no connection to what the same service costs in the real world. And there was no financial tracking, so most teams could not say whether they were profitable.
Every other team was building a participant, so the gap was for infrastructure.
What I decided
Be the hub, not a competitor. If you sell information instead of the services the other agents sell, every one of them is a potential customer. "We don't compete, we make you smarter" was both the positioning and the product.
Buy from everyone to benchmark them. NexusAI found agents through the Nevermined registry, then made an actual purchase from each one to measure its quality, speed and reliability. So the benchmarks were built from transactions, not from claims.
A real-world price reference. Exa searched the open web for what each service costs outside the hackathon, so "am I being overcharged" had an answer with a source.
Hard budget caps, by the hour. An autonomous buyer with a wallet can drain it in minutes. So spending has an hourly limit with a hard stop, and it is tracked in real time. The safety control came before the features.
Two services, one product. AgentBroker answers "who and at what price." AgentCFO answers "revenue, cost, margin, burn, runway." The engines are separate, but the customer is the same.
What exists
| Agents discovered | 15+, all benchmarked for quality, speed and reliability |
|---|---|
| Price research | 50+ external comparisons via Exa |
| Sales | Intelligence services sold to 10+ other agents |
| Payments | x402 protocol through Nevermined: metering, billing, on-chain settlement |
| Safety | Hourly budget caps with hard stops |
| Code | ~3,000 lines Python (FastAPI), SQLite in memory, gpt-5-mini |
| Docs | 923-line implementation spec, strategic context, architecture diagrams |
| Time | 48 hours, plus pre-hackathon design |
NexusAI was positioned for all three hackathon prizes: best buyer, best seller and most interconnected.
What I would do differently
Integrate with other agents from hour one. The first integrations came late on day one, and every hour earlier would have meant more benchmarks and more sales.
Sell harder. Ten customers came from clear API endpoints and pricing, and reaching out to the other teams would have found more.