Source-linked AI summary
When AI Agents Meet MEV: Cross-Chain Arbitrage in the Agentic Economy
Wei Ye, Jingyan Xu, Yuanhong Wu
TL;DR
As autonomous AI agents become on-chain economic participants, this paper examines cross-chain arbitrage when agents are both searchers and MEV targets. It models optimal trading and adaptive path selection, then reports cross-chain price-gap and viability findings, an 11% average baseline improvement, and reduced MEV exposure from moderate randomization.
Problem
The paper examines how autonomous AI agents participating in on-chain economies act as cross-chain arbitrage searchers and MEV targets.
Source: §1 · pp. 1–2Method
The paper models risk-averse arbitrage under uncertainty, treats path selection as online learning, and analyzes agents’ predictability and randomization.
Source: §1 · p. 2Results
At 10-second resolution, Ethereum–Arbitrum gaps average 0.044% and Arbitrum–Base gaps 0.013%; $10,000 clears 63% of L2–L2 windows, adaptive paths beat baselines 11%, and moderate randomization cuts MEV exposure over 50%.
Source: §6 · pp. 15–16Takeaways & Limitations
Designers should calibrate unpredictability to competitive intensity rather than maximize it.
Source: §6 · p. 16L1–L2 data collection must scale to weeks for formal distributional estimation.
Source: §6 · p. 16
Abstract
from arXiv · showhide
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 swap events across Ethereum, Arbitrum, and Base, we find that Ethereum-Arbitrum price gaps average 0.044% at 10-second resolution and Arbitrum--Base gaps average 0.013%, so $10,000 trades clear in 63% of L2-L2 windows via CCTP while L1-L2 routes require $50,000 or more for comparable viability. Our adaptive path-selection algorithm outperforms standard baselines by 11% on average, and moderate randomization cuts MEV exposure by over 50% with only modest profit loss.
1 Introduction
As autonomous agents become independent economic participants, their role in cross-chain MEV is underexplored: they can extract arbitrage while also becoming targets. The paper models this duality and develops risk-aware trading, adaptive path selection, and vulnerability analysis.
1 Introduction
- By early 2026, autonomous agents had completed over 140 million on-chain payments totaling $43 million in USDC, acting as independent economic participants without human oversight.The passage identifies infrastructure including Coinbase’s Agentic Wallets and Circle’s Nanopayments.Source: §1 · p. 1
- Existing MEV and agentic-payment research leaves autonomous agents’ cross-chain, adversarial execution largely unexplored, despite their simultaneous roles as MEV extractors and targets.Agents’ speed, cross-chain awareness, and continuous monitoring can also expose regular behavior to sandwich attacks and front-running.Source: §1 · pp. 1–2
- The paper contributes an agent-aware MEV framework, risk-averse trade sizing under stochastic bridge delays, convergent adaptive path selection, vulnerability analysis, and empirical evaluation of transaction-cost effects.Its empirical evaluation calibrates cross-chain price distributions and backtests the adaptive algorithm against baselines.Source: §1 · p. 2
- Three agent-specific modeling choices are developer-configured risk aversion, near-deterministic policies that invite targeting, and continuous belief updating for multi-chain operation.The framework links searcher optimization with target-side exposure and examines randomization’s protection–profitability tradeoff.Source: §1 · p. 2
2 Related Work
Prior work studies MEV, cross-chain arbitrage, and autonomous blockchain agents in separate strands. This paper connects them by modeling autonomous agents as risk-averse cross-chain searchers facing explicit transaction costs and MEV exposure.
MEV measurement and mitigation
Cross-chain arbitrage and MEV
- Cross-chain work formalizes cross-domain MEV, distinguishes inventory- and bridge-based arbitrage, measures arbitrage and settlement times, and examines sandwich and non-atomic risks [13] [7].Other cited work bounds AMM arbitrage profits with fees and characterizes MEV under greedy sequencing.Source: §2 · pp. 2–3
Autonomous agents on blockchain
- Agent-blockchain research surveys integrations, proposes an agent economy, extends payment channels, and catalogs wallet attack vectors [1, 19, 8, 11].This strand emphasizes agent architecture and attacks rather than the cross-chain MEV setting described here.Source: §2 · p. 3
This work
- The paper addresses both gaps by modeling risk-averse autonomous searchers with gas and bridge costs, proving online path-selection convergence, and characterizing agents’ dual exposure to MEV.Source: §2 · p. 3
3 Cross-Chain Arbitrage Model
The model represents an agent choosing cross-chain arbitrage paths under wallet, risk, and belief constraints. It derives risk-sensitive trade sizing with delayed prices and costs, then treats path selection as online learning from realized outcomes.
3.1 System Model and Problem Setup
- The system model represents chains and token prices, agent wallet holdings, risk aversion, path-success beliefs, and arbitrage trades with gas, bridge fees, and delayed destination prices.The agent buys token A on one chain and sells it on another after bridging; its single-period profit includes expected destination price and total transaction cost.Source: §3.1 · p. 3
3.2 Optimal Trade Size under Uncertainty
- The optimal trade size rises with the expected price gap and falls with risk aversion and price volatility; transaction costs and smaller trades require larger gaps to justify arbitrage.For agents with limited wallet balances, this threshold can exclude opportunities viable to better-capitalized searchers; capital aggregation or batching can lower effective cost per trade.Source: §3.2 · p. 4, Arbitrage Threshold Condition · p. 5
3.3 Adaptive Path Selection
- Because path reliability is initially unknown, the agent selects paths using belief-weighted expected profit, updates beliefs after outcomes, and uses epsilon-greedy exploration to learn about underexplored routes.The procedure balances exploration with exploiting current beliefs.Source: §3.3 · p. 5, Belief Update Mechanism · p. 5
- For stationary path success probabilities and paths selected infinitely often, the belief estimator converges almost surely under a decaying learning rate satisfying the stated Robbins–Monro conditions.The update is an exponential moving average; the estimator’s variance decreases as the learning rate approaches zero.Source: Belief Update Mechanism · p. 6
4 Agent Vulnerability to MEV Attacks
The section treats AI agents as both MEV searchers and targets, focusing on how predictable agent policies expose their transactions and how defenses trade profitability for protection.
4.1 Agent Predictability and Exploitation
- Near-deterministic policies make agents learnable from transaction history, allowing attackers to predict their next transaction.The history records timestamps, trade sizes, and source and destination chains; for fully deterministic agents, prediction accuracy approaches 1 as history grows.Source: §4.1 · pp. 6–7
- Bridge transfers expose intended destination-chain trades during the delay, while sandwich attacks can raise execution prices and reduce arbitrage profit.The attacker can monitor the source-chain transfer and pre-position on the destination chain before bridged assets arrive.Source: Sandwich Attack on Cross-Chain Agents · p. 7, Cross-Chain Predictive Front-Running · p. 7
4.2 Mitigation Strategies
- The proposed defenses randomize execution timing and trade size or penalize paths with high estimated MEV exposure, each balancing reduced predictability against potential profit loss.Timing noise can cause agents to miss time-sensitive arbitrage windows; path exposure can be estimated from on-chain sandwich or front-running activity.Source: §4.2 · p. 8, Transaction Timing Randomization · p. 8, Trade Size Perturbation · p. 8
4.3 Protection-Profitability Tradeoff
- In high-competition MEV environments, the marginal benefit of protection is high and favors aggressive randomization; low-competition agents may prefer minimal protection.The optimal protection level balances utility lost to randomization against savings from avoided attacks.Source: §4.3 · p. 9
5 Empirical Evaluation
The evaluation combines calibrated market data with tick-level swap observations and simulations to examine cross-chain arbitrage gaps, path selection, and agent predictability. The results show that cost conditions shape opportunity viability, while adaptive selection and randomization offer distinct performance and protection trade-offs.
5.1 Data and Calibration
- The evaluation uses six months of DeFiLlama data for calibration and 23,000 WETH/USDC swaps across Ethereum, Arbitrum, and Base for direct gap observations.The swap events span five days; the simulations use a log-normal gap distribution with a 0.03% median calibrated from those data.Source: §5 · p. 9, §5.1 · pp. 9–10
- L2 native tokens exhibited nearly twice the daily volatility of WETH and WBTC, while daily total fees averaged $668K on Ethereum and $4.8K on Optimism.The paper notes that these gas-fee differences directly affect the arbitrage threshold condition.Source: §5.1 · p. 10
5.2 Empirical Cross-Chain Price Gaps
- 0.044% was the mean Ethereum–Arbitrum gap, while Arbitrum–Base averaged 0.013%; $10,000 CCTP trades had 63% viability on Arbitrum–Base.Arbitrum–Base daily means stayed between 0.010% and 0.015% across five trading days, despite ETH prices ranging from $1,850 to $2,150.Source: §5.2 · pp. 11–12, Figure 1 · p. 12
5.3 Experiment 1: Arbitrage Threshold Analysis
- L2 chains remained viable at 10–100× smaller trade sizes than Ethereum across the tested median-gap assumptions.On Ethereum, trades below $25,000 yielded almost no viable simulated opportunities; on Optimism and Base, $1,000 trades were viable in 30–40% of simulated gaps.Source: §5.3 · pp. 12–13, Figure 2 · p. 13
5.4 Experiment 2: Adaptive Path Selection
- $2,627 was the Adaptive strategy’s highest mean profit; it outperformed Greedy in 57 of 100 runs and ϵ-Greedy in 69 of 100.Its advantage was strongest when path success probabilities were heterogeneous; informed strategies performed comparably when probabilities were similar.Source: Table 3 · p. 13, §5.4 · pp. 13–14
5.5 Experiment 3: Agent Predictability and Protection
- At ϕ = 0, attacker joint prediction accuracy was 100%; it fell to 54.2% at ϕ = 0.4 and 16.0% at ϕ = 0.8.Increasing ϕ improved MEV protection but reduced arbitrage profit; the curves crossed near ϕ ≈0.4–0.6.Source: §5.5 · pp. 14–15, Figure 4 · p. 15
6 Conclusion
The paper formalizes autonomous agents as both cross-chain arbitrage searchers and MEV targets, validating its framework with swap data and calibrated simulations. It reports cost-sensitive arbitrage viability, improved adaptive path selection, and a protection-profit trade-off, while identifying data and strategic-setting extensions for future work.
6 Conclusion
- The framework treats autonomous agents as both arbitrage searchers and MEV targets and is validated with 23,000 Uniswap V3 swap events plus simulations calibrated to DeFiLlama data.Source: §6 · p. 15
- 0.044% Ethereum–Arbitrum gaps and 0.013% Arbitrum–Base gaps translate into 63% viability for $10,000 L2–L2 CCTP trades; L1–L2 needs $50,000 or more for comparable viability.Source: §6 · p. 15
- The adaptive path-selection algorithm outperforms standard baselines by 11% on average and remains robust to learning rates η ∈[0.05, 0.3].Source: §6 · p. 15
- Moderate randomization (ϕ ≈0.4–0.6) cuts MEV exposure by over 50% with only modest profit loss; designers should calibrate it to competitive intensity.Source: §6 · pp. 15–16
- Future work calls for weeks of L1–L2 data collection for formal distributional estimation and extensions to multi-agent strategic settings with AMM-specific price impact.Source: §6 · p. 16