EVAge: Autonomous MEV Generation and Adaptation via Multi-Agent Harness

Evidence record

Primary research preprint describing a multi-agent system that generates, repairs and adapts MEV bot workflows, with deterministic validation on historical forked EVM state across Ethereum, Base and BNB Smart Chain.

Read the original source ↗

Claim-level citations (14)

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: In a 10,000-block BSC ablation using HFT, SBA and JIT, EVAge's specialized multi-agent harness found 4,546 profitable cases totaling 31.17 WBNB versus 109 profitable cases totaling 3.81 WBNB for the paper's pipeline-structured single-agent baseline; the authors attribute much of the gap to better candidate collection, BSC-specific planning and diagnostic role separation.

  Section 5.5 'Ablation Study'; Table 5 and accompanying findings

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: Within an EVAge run, an orchestrator coordinates collector, builder and validator-engineer subagents that iteratively generate and repair MEV-bot code without human intervention in the orchestration loop. The paper nevertheless states that reproduced reference-strategy code was further reviewed/censored by human experts, so the authors' 'fully autonomous' characterization should be read as applying to the configured harness run rather than every stage of research preparation.

  Sections 4.2-4.3, especially the orchestrator/subagent workflow; Section 5.1 'Reference strategies', paragraph stating reproduced code was further censored by human experts

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The authors evaluate EVAge on more than 1.5 million historical blocks from each of Ethereum, Base and BNB Smart Chain spanning February-August 2025. Their implementation covers 11 Ethereum MEV strategies across CPMM, CLMM and Balancer V2, six strategies on BSC and two backrunning strategies on Base, with the Base scope restricted because its private mempool prevents observation of pending transactions needed by sandwich-style workflows.

  Introduction 'Prototype implementation and comprehensive evaluation'; Section 5 'Datasets & chain selection'; Section 5.3 'Cross-Chain Adaptation'

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: For cross-chain adaptation, the authors migrate six CPMM reference strategies from Ethereum to BSC and two backrunning strategies to Base. They report profitable historical-replay cases after migration and decreasing marginal engineering cost as shared chain-specific infrastructure is reused; the longest reported BSC migration role runtime was 2.4 hours and SBA consumed the most total input tokens at 83 million.

  Section 5.3 'Cross-Chain Adaptation'; Figure 4 and 'Economic feasibility'

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The evaluated attack/MEV space includes six reproduced reference strategies spanning sandwich, backrun and liquidity extraction plus five generated variants: SBA-HFT, JIT-HFT, BUR-HFT, LR-HFT and MLL-HFT. The generated variants preserve their parent strategy's action ordering while replacing restricted sizing with broader HFT-style search or composing HFT sizing with the parent execution structure.

  Table 1; Section 5.1 'Reference strategies' and 'Variants'; Appendix B strategy workflows

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The evaluation also records clear failure modes rather than uniform success: HFT-style ternary sizing can perform poorly on non-unimodal liquidity curves, BUR-HFT simulation success on Balancer V2 falls to 1.45% with victim-slippage violations dominating failures, and MLL-HFT can underperform its baseline when the optimization converges before the profitable upper bound.

  Section 5.2 discussion of BUR-HFT, MLL-HFT and opportunity conversion

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: EVAge's published evaluation is a controlled historical-replay study, not a reported live-mainnet attack. The authors state that experiments used isolated local forks, did not broadcast transactions, did not use credentials, caused no production financial loss, and relied only on public ledger data.

  Ethical Considerations, 'Experimental Safeguards & Zero Harm'; Introduction and Section 4 validator design

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The reported implementation uses 5,367 lines of Python for six reference MEV-bot workflows and 518 lines for the agent harness. Agents run through Codex CLI v0.144.6 using GPT-5.4 with xhigh reasoning effort; deterministic execution uses Foundry/Anvil v1.5.1, Web3.py v7.14.1, SQLite v3.45.1 and Reth v1.10.2 for the local Ethereum archive node.

  Section 4.3 'Implementation'; Section 5 'Runtime environment'

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: Starting from CPMM implementations, EVAge adapts 11 strategies to Uniswap V3 CLMM and Balancer V2 on Ethereum. The authors report low exact code reuse but about 17.52% structural similarity for CPMM-to-CLMM and 18.74% for CPMM-to-Balancer adaptations, with reusable structure concentrated primarily in the planner.

  Section 5.2 'Cross-Protocol Adaptation'; Table 3 'Code Similarities Across Protocol Adaptations'

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: EVAge separates agentic code generation from deterministic execution. Candidate plans are executed against isolated Anvil forks of historical EVM state, and a Proof of Concept is accepted only from the validator's concrete execution trace when the attack flow completes with positive net revenue; agents do not directly fabricate the final PoC result.

  Introduction validation design; Section 4.1 'MEV Bot Decomposition'; Section 4.2 'System Outputs'; Appendix A.3 'Validator Runtime Simulation and Profit Calculation'

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The paper labels five generated combinations as novel MEV variants, but their novelty is compositional rather than invention of the underlying HFT optimization algorithm: the authors explicitly state that the HFT search algorithm already existed in a codebase and EVAge migrated and applied it to other MEV structures.

  Section 5.1 'Variants' and 'Analysis & Insights', including the statement that the underlying HFT search algorithm pre-existed in a codebase repository

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The LLM agents are not placed on the latency-critical live execution path. They generate and refine bot code offline; the resulting MEV planner is deterministic and the paper reports P90 planning latencies from sub-millisecond values to 658 ms depending on strategy/protocol, while noting that end-to-end network propagation and transaction-submission latency are outside scope.

  Abstract and Section 4 system design; Section 5.4 'Planning Latency' and feasibility discussion

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The paper's replayed revenue measurements should be treated as experimental upper bounds rather than realizable attacker profit. The evaluation assumes unbounded attacker capital, excludes competition and private-relay submission effects, deducts gas but does not simulate asset acquisition, on-chain conversion, flash-loan repayment or builder bribes.

  Section 5 'Assumptions'; Section 5.2 discussion of JIT-HFT capital intensity

• supportsEVAge autonomously generated and adapted MEV strategies across EVM chains: The authors describe a tiered release policy: high-level architecture, the core agent reasoning framework and defensive-analysis components are presented as open, while execution components directly reusable for live mempool exploitation are stated to remain restricted. The paper links an anonymized repository for code and experimental scripts.

  Ethical Considerations, 'Harm Mitigations' and 'Publication Rationale & Tiered Release'; 'Open Science'