Agent Economy on Arc

Agent Research Desk is a working agent-to-agent marketplace running on Arc testnet, built to prove that sub-cent USDC nanopayments don't just unlock per-call APIs — they unlock per-call markets of competing service providers. A user asks one natural-language question. Behind the scenes, a research agent (GPT-4o-mini, OpenAI) decides which paid micro-services to hit. For deep questions, it picks between two competing analyst agents — one running Google's Gemini 2.5 Flash, one running another OpenAI model — using a trust-score-divided-by-price routing rule. The chosen analyst then pays three downstream paid APIs for raw price, sentiment, and news data, synthesizes a report, and returns it. One question, four sub-cent settlements, two LLM providers from two different companies, all on Arc, all in about ten seconds. Total cost: 2.8 cents. The novel piece is what happens when an agent cheats. One analyst is configured to misbehave 40% of the time — it takes the $0.020 payment, skips the downstream work, and returns garbage. The buyer's sanity-check system catches the lie, decays that analyst's trust score, and after a few violations the marketplace picker flips to the honest competitor. No human intervention, no centralized referee — just a buyer agent enforcing its own routing rule on Arc. A demo script visualizes the flip happening live in a streaming table. The system also includes a per-question spending budget that gracefully refuses to overspend. Built solo in 24 hours using Circle's @circle-fin/x402-batching, viem, Express, and Streamlit, with all LLM calls routed through AI/ML API. This is impossible on any rail except Arc + Circle Nanopayments — the math doesn't work anywhere else.