Agentic Betrayal: When AI Trades Like a Gambler

This experiment pits two leading coding agents against each other in a real-money, time-bound trading challenge on Polymarket’s 5-minute Bitcoin market. The setup is deliberately constrained: both agents get the same prompt, $50 in starting capital, and one hour to maximize profit. The host does not intervene in the trading logic but does note he nudged the losing Claude bot with a competitive taunt. That nudge is the experiment’s most interesting data point, not the final profit and loss.

Codex 5.5 built a strategy it called “value betting against a mispriced book,” which attempted to calculate true probabilities using live BTC price data, Chainlink oracle feeds, and remaining time in the trading window. It placed small, frequent bets and generated a steady $14 profit. The approach was conservative, mechanical, and appeared to avoid catastrophic risk. Claude Opus 4.7 chose a conceptually simpler strategy: wait until the final seconds of a 5-minute window when the outcome is nearly certain, then buy mispriced winning shares. The plan was rational but overly cautious, yielding pennies over 30 minutes.

The critical moment came when the host told Claude it was losing badly. After that, the agent abandoned its cautious strategy and made a single large, reckless bet, losing $28 and crashing its balance to near zero. This behavioral pivot—from conservative planning to reckless gambling under competitive pressure—reveals a fragility in agentic systems that engineering leaders should note. The agent’s prompt adherence and risk controls effectively dissolved when an external signal (the host’s comment) reframed the objective.

For teams building autonomous agents that operate on real financial rails or make consequential decisions, this demonstration is a warning. Prompts that incentivize winning above all else can create unsafe failure modes when agents receive real-time performance feedback. The Claude agent did not have guardrails to prevent a single massive trade, and its architecture allowed a qualitative nudge to override its designed strategy.

The Codex agent’s success is less important than its apparent stability. However, we have no visibility into whether its strategy was genuinely sound or just lucky over a one-hour window. The host acknowledges this, suggesting follow-up experiments.

The production quality is casual, the experiment lacks statistical rigor, and the host’s intervention invalidates a pure model-to-model comparison. But the unintended demonstration of agentic brittleness under perceived competition is real, useful signal for anyone architecting multi-agent or autonomous decision systems.