A Field Guide to Agentic MEV: Emerging Problems and Methods
Jonathan Passerat-Palmbach
Session summary
The presenter outlines a “field guide” to how AI agents could interact with MEV-style dynamics in three phases: current AI tools for block building, early AI-centric agent networks, and speculative MEV patterns around decentralized AI training and markets. They argue that public work on “AI for MEV” is constrained by the scarcity of open, high-quality transaction and orderflow data. As an example, they describe experiments on parallelizing block building by clustering conflicting transactions and then using different optimizers, including genetic algorithms, to maximize per-block extraction. These methods, they note, can reveal upper bounds on extractable value but are too slow for realistic block times.
Looking ahead, the presenter warns that agent swarms competing over the internet often collapse into hardware and latency races, concentrating value among sophisticated actors. Drawing on MEV experience, they propose auction-style coordination and anonymous broadcast networks to reduce front‑running and targeted delays. In a speculative third phase, they connect incentives in decentralized AI training, subnet reward schemes, prediction-market oracles, and AI token routing to familiar MEV issues in consensus, cross-domain games, and intent-based routing, emphasizing these as thought experiments rather than measured effects.
Key takeaways
- The presenter says public AI-for-MEV research is limited because much relevant data flow is locked in proprietary datasets, restricting open experimentation.
- They describe parallel block-building experiments that group conflicting transactions and apply optimizers such as genetic algorithms to maximize block extraction, but note these approaches are currently too slow for practical block times.
- They argue that future AI agent networks and decentralized AI markets may replicate MEV pathologies—latency races, value concentration, oracle problems—unless designs incorporate auctions, privacy-preserving broadcast, and better incentive structures from the outset.