AI Crypto Trading Agents: Hype vs Reality in the Wild West of Markets

The hype around AI crypto trading agents has reached new heights, but what can they actually do? In reality, these agents are not as autonomous as they seem. According to a recent Bank of England report, most live stacks are 'AI-assisted' rather than 'AI-directed,' with the human remaining the portfolio manager.

So, what exactly can AI crypto trading agents do? They can automate research digests and data labeling, run fully autonomous multi-venue portfolios without human review, generate and refactor backtest code fast, and execute small orders in stable liquidity windows. However, they usually cannot do these tasks reliably yet.

The Bank of England's report highlighted that firms are mostly using AI for lower-risk work such as research, coding support, and surveillance rather than letting it fully run portfolios. This is because crypto markets are hostile to naive automation, with latency games, fee drag, fragmented liquidity, and sudden regime shifts punishing models that look great on static charts.

When an AI trading agent does go live, the results can be disastrous. A recent audit found that costs crushed pretty pictures, with frictions eating the edge of a model that leaned on frequent switching or chased tiny signals. Overfitting to calm regimes also proved to be a major issue, as agents trained on long stretches of range-bound chop often blew up on regime flips.

So, how can you build a realistic agent stack? Start by defining the agent's remit and wiring data with redundancy. Backtest with realistic costs and partial fills, not fantasy fills, and dry run in paper or shadow mode behind a human for 2 to 4 weeks. Finally, start live tiny, track realized slippage by venue and by hour, and set kill-switches.