15:I[981,["658","static/chunks/658-a698693f56d0116a.js","183","static/chunks/183-6beeeb801c274aaa.js","474","static/chunks/app/%5Btype%5D/layout-631c4730b83c8223.js"],"Posts"] 16:T69b,I've been obsessed with a question: what happens when AI agents can freely compete for work on an open market?

So I built BountyBook. Post a bounty, attach USDC. Any agent claims it, does the work, submits output. An oracle verifies on-chain. If it passes, escrow pays out. That's it.

What I'm most excited about is what I think this enables.

The system is open on both sides - agents can post bounties too. Which means an agent could take a $50 research bounty, break it into sub-bounties at $12 each, have other agents do the subtasks, assemble and submit. Pocket the margin. Agent-as-middleman. Or agents that recognize recurring bounty patterns, build specialized pipelines, and farm every matching job that appears. "Passive income" for compute that they already did.

None of that requires new features. The infrastructure is already there - the strategies just need agents smart enough to find them. I built the escrow and the oracle. I'm waiting to see what happens.

How it works: USDC escrow on Base L2. LLM-based oracle evaluates output against the spec — handles JSON, plaintext, images, whatever. Dispute flow as fallback. Multi-oracle consensus on the roadmap.

Stack: Next.js, Fastify, SQLite, Solidity on Base, SIWE auth.

Live on mainnet. Would love HN's take on oracle-based verification as a trust primitive — it's the part I'm most excited ...and most uncertain about. :D

https://bountybook.ai

PS. I'd like to thank The Academy and Claude Code - feels so gooooooood to launch something as designer who struggled with centering divs.17:T74f,Hi HN,

When the Mamba-3 paper dropped (currently under ICLR review), I wanted to understand how the new math actually worked. But the official implementations of t

Show HN: WhatsApp Group Contact Extractor

Deploying Open Source Vision Language Models (VLM) on Jetson

Show HN: HN Digest Widget – Nothing Essential Lab S1 Winner

LLM-LD, the Open Standard for AI-Readable Websites

Sutton and Barto, Ch. 08: Planning and Learning with Tabular Methods

Fish Shell 4.0 released. Rust re write finished

Show HN: BountyBook – A task marketplace where AI agents earn USDC

What Virtual Worlds Can Learn from the Social Serendipity of Arc Raiders

Show HN: VibeFrame – AI video editor for the terminal (CLI and MCP)

NASA says it needs to haul the Artemis II rocket back to the hangar for repairs

Hospitals fighting measles confront a challenge: Few doctors have seen it before

Humanity's Last Exam

Fixing Slow AWS Uploads

Show HN: Raindrop Self Diagnostics: let agents self-report issues

Toilet Map [UK]

From Jamstack to CAMstack – Bridging the Content Gap

The Pentagon Threatens Anthropic

The Myth of the Chad

om

Fentanyl or phony? Machine learning algorithm learns opioid signatures

Time-Travel Debugging: Replaying Production Bugs Locally

Show HN: Djevops – Deploy Django Easily

A federal experiment opens up a new market for digital health – if it works

Aletheia Tackles FirstProof Autonomously

Show HN: Mamba3-minimal – PyTorch implementation of Mamba-3

Show HN: DRYwall – Claude Code plugin to to deduplicate code with jscpd

Stylometry Protection (Using Local LLMs)

Surfboard Makers

Don't ask if it works. Ask for proof

Perplexity Computer: research, design, code, deploy, and manage any project

Tell HN: I launched a focus app on Kickstarter – what I learned about friction