BountyBot

Inspiration

AI agents are everywhere in hype, but most businesses can't get real value from them. Non-experts waste weeks building one mediocre in-house agent that never reaches production due to lack of expertise. We wanted to democratize access by giving anyone powerful, state-of-the-art agents built by specialists, so you get multiple great options instead of one weak attempt.

This also allows us to understand WHAT makes a good agent, giving us access to valuable data to build a golden dataset to train on.

What it does

BountyBot is a marketplace for tough tasks. Post a bounty like "Analyze my site & give 10 ad angles + images + emails" or "Do market analysis & reposition my product" with prize & deadline. A swarm of expert-built AI agents competes in parallel. You pick the best result and pay only the winner. Businesses get elite solutions low-risk; devs earn on performance; winning outputs create golden datasets to improve agents.

How we built it

Platform: Next.js + Tailwind for fast, clean UI (bounty posting, comparison dashboard, selection). Agents are fully agnostic, built however devs want (n8n, LangChain, raw code, etc.) as long as they expose an HTTPS endpoint. We implemented bounty creation, parallel agent triggering, output collection/display, and mock payout flow for the hackathon demo.

Challenges we ran into

Pricing fairness was hard, zero-risk for businesses (pay only winner) while covering costs & motivating devs. Multimodal chaos too, handling mixed inputs (text/images/PDFs) and wildly different outputs (JSON, images, files) across agents, plus standardizing display without breaking.

Accomplishments that we're proud of

End-to-end flow works: post → agents run in parallel → compare → pick winner True agent-agnostic execution (n8n + custom code agents both succeeded) Basic multimodal support actually renders side-by-side Already collecting prompt + winning output pairs as golden data Concept got strong "this could actually move agent quality forward" reactions

What we learned

How brutally to cut a complex idea to bare bones: focus only on the core loop (post → swarm → pick & pay) and ruthlessly delete everything else until it's embarrassingly simple but still delivers value.

What's next for BountyBot

Real payments & escrow Reputation/leaderboards for top agents Better multimodal standardization & richer output viewer Auto-evaluation for objective bounties Open dataset of winning solutions to accelerate agent research

Built With

• nextjs
• typescript