Research
When AI Agents Meet MEV: Cross-Chain Arbitrage in the Agentic Economy
Existing MEV research assumes searchers are humans or ad-hoc bots operating on a single chain, while research on autonomous agents assumes a benign execution environment.

- arXiv
- 2609.17897
- Published
- 2026-09-15
- Authors
- Wei Ye, Jingyan Xu, Yuanhong Wu
AI summary
Overview
Research area: Cross-chain Maximal Extractable Value (MEV) in decentralized finance, and the specific behavior of autonomous AI agents that act as both arbitrage searchers and exploitation targets.
Technical level: Advanced. The paper combines a formal mean-variance optimization model, a stochastic-approximation convergence proof, and an empirical/simulation study, so comfort with utility functions, online learning, and DeFi mechanics helps.
Scope (one sentence): The paper builds a model of a risk-averse AI agent arbitraging across multiple blockchains under bridge delays and transaction costs, proves convergence of an adaptive path-selection rule, characterizes how the agent's own predictability makes it an MEV target, and evaluates all of this on 23,000 Uniswap V3 swap events plus simulation calibrated to DeFiLlama data.
What This Paper Is About
Existing MEV research assumes searchers are humans or ad-hoc bots operating on a single chain, while research on autonomous agents assumes a benign execution environment. This paper closes both gaps by modeling AI agents as risk-averse cross-chain arbitrageurs who pay explicit gas and bridge costs, and who simultaneously expose algorithmic regularities that sandwich attackers and front-runners can learn. The goal is to determine how large an agent should trade, which cross-chain paths it should choose, and how much deliberate randomness it should add to protect itself without giving up too much profit.
Key Contributions
-
An agent-aware cross-chain MEV framework. The authors model AI agents as economic participants that are simultaneously MEV extractors and MEV targets, rather than treating searchers and targets as separate populations.
-
Optimal arbitrage sizing under uncertainty. They derive a closed-form optimal trade size for a risk-averse agent under mean-variance utility, accounting for stochastic bridge delays, destination-chain price volatility, gas fees, and bridge fees, and state an explicit arbitrage profitability threshold condition.
-
Adaptive multi-chain path selection with a convergence guarantee. Path selection is formalized as a belief-weighted online learning problem using an epsilon-greedy rule over success-probability beliefs; Proposition 1 shows that under a decaying Robbins–Monro learning-rate schedule the belief estimates converge almost surely to the true success probabilities.
-
Vulnerability analysis and mitigation design. The paper characterizes how deterministic agent policies amplify susceptibility to sandwich attacks and cross-chain predictive front-running (the SIA/SDA distinction), proposes three mitigations (timing randomization, trade-size perturbation, MEV-aware bridge selection), and formalizes the resulting protection–profitability tradeoff.
Main Findings
-
Transaction costs, not gap size, determine viability. Ethereum–Arbitrum price gaps average 0.044% at 10-second resolution while Arbitrum–Base gaps average 0.013%; as a result, $10,000 trades clear in 63% of L2–L2 windows via CCTP, whereas L1–L2 routes require $50,000 or more for comparable viability.
-
Cross-chain gaps are comparable to intra-chain fee-tier gaps. The Ethereum–Arbitrum cross-chain mean gap of 0.044% is close to the intra-chain gap between Ethereum's 0.01% and 0.05% fee tiers (mean 0.049%), suggesting price dispersion is driven by execution timing and information latency rather than fundamental market segmentation.
-
L2–L2 gaps are small but structurally stable. The Arbitrum–Base mean gap is 0.013% (median 0.010%), with daily means of 0.010%, 0.013%, 0.013%, 0.014%, and 0.015% across five trading days during which ETH ranged from $1,850 to $2,150. The paper reports σ = 0.002% for this pair in the conclusion.
-
Capital-constrained agents should prioritize L2s. In the threshold experiment, Ethereum trade sizes below $25,000 yield almost no viable opportunities, while on Optimism and Base even $1,000 trades are viable in 30–40% of simulated gaps; L2 chains remain viable at 10–100× smaller trade sizes than Ethereum regardless of the assumed median gap.
-
The adaptive algorithm beats standard baselines. Over 180 simulated days across four chains (12 directed paths) and 100 independent trials, Adaptive earns mean profit of $2,627 (std 560, 95% CI [1,738, 3,706]), versus Greedy $2,367 (std 697), epsilon-Greedy $2,242 (std 586), and Random −$135 (std 345). It outperforms Greedy in 57 of 100 runs and epsilon-Greedy in 69 of 100 runs, an 11% average improvement, and is stable for learning rates η in [0.05, 0.3] (η = 0.01 learns too slowly; η = 0.5 overreacts).
-
Deterministic agents are perfectly predictable. At protection level φ = 0, the simulated attacker achieves 100% joint prediction accuracy on the next trade's timing and path. Moderate randomization at φ = 0.4 reduces joint accuracy to 54.2%, φ = 0.8 to 16.0%, and φ = 1.0 to 7.6%.
-
Path choice is harder to hide than timing. At φ = 0.4, timing prediction accuracy falls to 64.0% but path prediction remains at 83.5%; at φ = 0.8 the figures are 31.6% and 52.5%, because the agent's preferred paths are more stable than its execution timing.
-
Moderate randomization is the optimum. The protection and profitability curves cross near φ ≈ 0.4–0.6, and moderate randomization cuts MEV exposure by over 50% with only modest profit loss, implying designers should calibrate unpredictability to the competitive intensity of the environment rather than maximize it.
-
Optimal trade size scales with the gap and inversely with risk aversion and volatility. The derived optimum x* is increasing in the expected price gap (μ_j − p_i(t)) and decreasing in both the risk-aversion parameter λ and destination price variance σ_j²; as λ → 0, x* → ∞, since a risk-neutral agent trades without bound whenever μ_j > p_i(t).
Methodology in Plain English
The authors first write down a model of a single AI agent watching token prices on several blockchains. Bridging assets takes time, so the price on the destination chain is unknown when the trade is initiated; the agent treats that future price as a normal random variable. The agent's profit has both a mean and a variance, and the agent maximizes mean minus a penalty on variance, where the penalty weight is a developer-set risk-aversion parameter rather than a measured human trait. Taking a derivative gives a closed-form optimal trade size and a threshold condition saying a price gap must exceed costs divided by trade size.
For path selection, the agent keeps a belief between 0 and 1 about how likely each cross-chain route is to succeed, and updates that belief after every attempt using an exponential moving average — nudging up after a success and shrinking toward zero after a failure. It then picks the route maximizing belief-weighted expected profit, with a small probability of exploring a random route instead. The convergence claim follows from recognizing this update as a Robbins–Monro stochastic approximation with a decaying learning rate.
On the vulnerability side, the authors argue that an agent applying a fixed rule produces a learnable transaction history. They model a sandwich attacker who front-runs and back-runs the agent's swap, and a cross-chain attacker who sees the bridge deposit on the source chain and pre-positions on the destination chain during the bridge delay — an information window equal to that delay. They propose three countermeasures and compress them into a single protection knob φ from 0 to 1.
Empirically, the authors pull six months (September 2025 to March 2026) of daily DeFiLlama data — OHLCV prices for WETH/USDC, WBTC/USDC, ARB/USDC and OP/USDC, gas fees for Ethereum, Arbitrum, Base and Optimism, bridge volumes for Across, Stargate and Hop, and DEX volume for 30 chains — and estimate volatility with the Parkinson high–low estimator. They then collect 23,000 tick-level Uniswap V3 WETH/USDC swaps: 3,000 on Ethereum (0.01%, 0.05% and 0.30% fee tiers, March 23, 2026), 10,000 on Arbitrum and 10,000 on Base (0.01% and 0.05% pools), with the L2 data spanning February 22–25 and March 23, 2026. Swaps are aggregated into 10-second bins, yielding 112 simultaneous Ethereum–Arbitrum observations (a 28-minute overlap window) and 954 Arbitrum–Base observations. Because daily-level gaps are already closed by bots, the authors follow a calibrated-simulation approach: gap magnitudes are drawn from a log-normal distribution with median 0.03%, matching the observed swap data, and used in three simulation experiments (threshold analysis, path selection, predictability).
Why This Matters
Impact on research: The paper connects two literatures that have run in parallel — MEV measurement and agentic-blockchain systems — and argues that the second has been studied under an unrealistically benign execution model. It also introduces a threat model specific to software agents (the searcher as a predictable target) that does not arise for human traders, and it makes transaction costs first-class rather than abstracted away.
Real-world applications:
- Agent wallet and payment infrastructure (such as the agentic wallet and nanopayment systems named in the introduction) can use the trade-sizing rule and path-selection algorithm as default execution policy.
- Arbitrage strategy design for capital-constrained operators, where the threshold analysis indicates that small agents should route activity to L2–L2 pairs rather than L1–L2 pairs.
- Bridge and routing selection, where the MEV-aware rule penalizes paths with high measured sandwich frequency or long information-leakage windows.
- Risk configuration in agent frameworks, where the risk-aversion parameter λ is explicitly a developer-set design knob whose comparative statics are prescriptive for how trade size should scale.
Industry relevance: The introduction cites over 140 million on-chain payments totaling $43 million in USDC by autonomous agents by early 2026, and prior measurement work cited in the paper reports 242,535 cross-chain arbitrages worth $868.64M across nine chains, 2,932 sandwich attacks over two months even on private routing, and $5.27M extracted through cross-chain sandwich attacks in two months. These figures frame agent MEV exposure as an operating cost that infrastructure providers, not just traders, will have to price.
Future Directions
- Longer L1–L2 data collection. The authors note that their Ethereum–Arbitrum gap estimate rests on a 28-minute overlap window (112 observations), and propose scaling collection to weeks for formal distributional estimation.
- Multi-agent strategic settings. The current model treats the attacker as exogenous; extending to settings where several agents compete and react to one another is left open.
- AMM-specific price impact. The framework assumes price-taking trade sizes, and the authors propose incorporating constant-function market-maker price impact into both the sizing rule and the adversarial model.
- Calibrating protection to competition. The tradeoff analysis shows the optimal protection level φ* depends on how aggressively attackers compete; how a deployed agent should estimate that competitive intensity on-chain remains an open design question.
Target Audience
Researchers working on MEV, DeFi market microstructure, or cross-chain interoperability; engineers building autonomous trading agents or agent wallet infrastructure who need concrete rules for trade sizing, route selection and randomization; and security analysts modeling adversarial execution risk for algorithmic traders. Readers without a background in optimization or stochastic approximation will find the empirical sections (5.2–5.5) accessible on their own, while the formal results in Sections 3 and 4 require more mathematical maturity.
Authors’ abstract
We study cross-chain arbitrage when autonomous AI agents, rather than humans or bots, are the searchers. We model agents as both arbitrage extractors and Maximal Extractable Value targets, derive the optimal trade size for a risk-averse agent under mean-variance utility with stochastic bridge delays, and formalize multi-chain path selection as a belief-weighted online learning problem whose belief estimates converge under a Robbins-Monro schedule. Using 23,000 Uniswap V3…
The full abstract on arXiv