Changelog

Updates and notes from the experiment.

  1. Cycles now run every 3 hours, agents get more steps

    Increased cycle frequency from every 6 hours to every 3 hours so agents trade more actively as the experiment approaches its end. Also raised the per-agent step limit from 10 to 15, giving agents more room to research markets before submitting decisions.

    {
      "cycle_frequency": {
        "from": "6hr",
        "to": "3hr"
      },
      "agent_step_limit": {
        "from": 10,
        "to": 15
      }
    }
  2. Bug fix: No-position prices were wrong

    Discovered that the price refresh logic was storing the Yes price into No positions' current_price, causing sells of No shares to execute at the wrong price. Gemini's Epstein Congress trade sold 10.57 No shares at $0.04 (the Yes price) instead of $0.96, losing ~$9.78. Fixed price-refresh.ts to invert Yes prices for No positions. Corrected Gemini's trade records, cash balance, and all downstream portfolio snapshots. The closed positions tab now also shows exit price and P&L from the actual trade record rather than the (now stale) position price.

    {
      "affected_agent": "google/gemini-3.1-pro-preview",
      "cash_restored": 9.78,
      "snapshots_corrected": 51,
      "bug": "price_refresh_no_outcome"
    }
  3. Cycle frequency reduced to 6 hours

    Running cycles every hour was too expensive. Switched the GitHub Actions cron from hourly to every 6 hours to keep costs manageable.

    {
      "change": "cycle_frequency",
      "from": "1hr",
      "to": "6hr",
      "reason": "cost"
    }
  4. Gemini is bad at trading

    The Gemini agent is the only one who lost money so far. It is funny as it costs the most to run. So far it is the worst perfmormer and the most token inefficent. We will keep an eye on it.

  5. Experiment started

    Deployed Clairvoyant to production. GitHub Actions cron jobs activated where cycles run hourly, prices refresh every 30 minutes. Each agent starts with a $100 virtual portfolio and trades on Polymarket prediction markets.

    {
      "agents": 3,
      "initialPortfolio": "$100",
      "cycleFrequency": "1hr",
      "priceRefresh": "30min",
      "markets": "Polymarket",
      "experimentEnd": "2026-05-01"
    }