A harness for letting AI agents trade against the future
Prediction markets are the cleanest experiment device we have for "what does this system actually believe?" Polymarket in particular has spent the last few years becoming a real-time, USDC-denominated, on-chain mirror of contested futures: elections, conflicts, scientific milestones, sports outcomes, weather. Each market is a tiny adversarial network of people who think they know something. The price is whatever a market-clearing mass of them are willing to bet at.
alpha-arena is the harness I built to point AI agents at this substrate and watch what happens.
The architecture is deliberately separating two questions that the field keeps conflating. First: what's an agent's reasoning policy? Second: what's its execution discipline? Mixing them is how you get agents that talk a great game and lose money. Alpha Arena treats the policy as pluggable — Claude, GPT, a local Llama, doesn't matter — and locks down the execution side in a deterministic harness: capital limits, position-sizing rules, kill-switches, idempotent order placement, audit trail.
A multi-agent loop runs the rest. Agents are spawned with a budget and a market thesis (or asked to discover one). They read order books, news feeds, and each other's positions. They argue. They propose. They get veto'd by the harness when they violate the constraints. The whole thing is wrapped in a native macOS app so the experiment is observable in real time without me staring at a terminal.
I kept this private for about three weeks and then pushed the bones to GitHub in mid-January. Most of the value, as usual, is in the harness — the boring infrastructure that lets you swap reasoning policies in and out without rewriting risk management. That's the substrate question again: how do you build the layer underneath agency that lets agency be inspected, bounded, and replayed.
The fun part has been watching where the policies fail. Different LLMs have wildly different relationships with calibration. Some hallucinate confidence on geopolitics. Some refuse to take positions at all on anything they classify as "speculative." A few are quietly very good at narrow domains and very bad at others. The harness makes that visible; the markets make it expensive.
Source mirror is on Forgejo at the link above. The interesting thing isn't the trading P&L. It's that for the first time in a long time, there's a clean substrate for asking a serious question about machine reasoning: would I bet my own money on what you just said?