A Field Guide to Agentic MEV: Emerging Problems and Methods
Jonathan Passerat-Palmbach
Session summary
The presenter outlines three ways AI and MEV (maximal extractable value) interact: using AI to build better blocks, designing MEV-aware AI agents, and anticipating MEV-like effects in emerging AI-native networks. He describes a “parallel block building” approach that partitions transactions into conflict groups, then experiments with search techniques such as genetic algorithms and reinforcement-learning-style methods over the EVM as a simulator to improve block ordering and value. When combined with auto-research loops, he argues that single-objective optimization on block value can autonomously rediscover strategies like sandwiching, motivating multi-objective reward functions and explicit “do no harm” constraints.
He then explores how distributed auto-research swarms and AI networks can reproduce familiar MEV dynamics, including latency races and rich-get-richer effects for actors with better hardware or stake. Using examples such as Bittensor-style subnets, proof-of-learning proposals, prediction-market oracle agents, and AI routers, he highlights risks of free-riding, manipulation of activity metrics, and front-running of predictions. The presenter suggests reusing MEV mechanism-design tools—auctions, privacy primitives, anonymous communication, and better coordination protocols—and mentions constrained data-access setups and formalized safety policies as promising research directions.
Key takeaways
- The presenter reports that parallel block building combined with search methods like genetic algorithms and EVM-based reinforcement learning can improve block value, but also shows how auto-research loops need explicit multi-objective and “do no harm” constraints.
- He argues that distributed auto-research swarms and loosely coordinated AI agents are prone to latency races and rich-get-richer dynamics, especially when some participants have superior GPUs or infrastructure.
- Using AI-native networks and oracle systems as examples, he highlights MEV-like incentives for validators to free-ride, manipulate activity, or front-run predictions, and suggests adapting existing MEV tools such as auctions and pre-trade privacy to mitigate these behaviors.