Research
EVAGE: Autonomous MEV Generation and Adaptation via Multi-Agent Harness
MEV capture on blockchains is dominated by well-resourced quantitative teams because both designing a strategy and implementing it for each new protocol or chain require slow, manual expert…

- arXiv
- 2609.27424
- Published
- 2026-09-23
- Authors
- Yan Wen, Zichun Cai, Iliya Mirzaei, Xiaohua Cai, Mohammad Javad Amiri, Haoxian Chen, Chenyuan Wu
AI summary
Overview
Research area: Blockchain economics and security — specifically Maximal Extractable Value (MEV) extraction on EVM-compatible chains — combined with LLM-based multi-agent code generation.
Technical level: Advanced. The paper assumes familiarity with AMM mathematics (CPMM, CLMM ticks, Balancer weighted pools), mempool and bundle mechanics, forked-EVM simulation, and LLM agent orchestration.
Scope (one sentence): The paper presents EVAge, a multi-agent LLM harness that autonomously writes, repairs, and validates MEV bot code offline — discovering new strategy variants, porting strategies between AMM protocols, and migrating them across Ethereum, Base, and BNB Smart Chain — evaluated on over 1.5 million historical blocks per chain.
What This Paper Is About
MEV capture on blockchains is dominated by well-resourced quantitative teams because both designing a strategy and implementing it for each new protocol or chain require slow, manual expert engineering. EVAge attacks this bottleneck by using a coordinated set of LLM agents to write and iteratively repair the actual MEV bot code ahead of time, so that no model inference sits on the latency-critical execution path. The goal is to let a user state a high-level objective — find a new variant, adapt to a new AMM, or port to a new chain — and have the system produce a deterministic, simulation-verified Proof-of-Concept that earns positive net revenue.
Key Contributions
-
Autonomous MEV generation. EVAge is presented as the first agentic system to generate MEV bots end to end without human intervention, synthesizing strategy logic from existing literature, implementing it, and adapting it to new protocols and chains — validated on over 1.5 million historical blocks.
-
Hierarchical multi-agent harness. The MEV bot is decomposed into three stages — Data Collector, MEV Planner, and Validator — each owned by a dedicated subagent (collector engineer, builder, validator engineer) under an orchestrator that routes control and diagnostic feedback in a closed loop. This decomposition limits each agent's context, reduces hallucination, and keeps the reasoning layer separate from deterministic execution.
-
Cross-protocol and cross-chain adaptation. Because protocol- and chain-specific changes are largely confined to the collector and validator, the core planner logic is reused. Adapted implementations show 18.13% average structural code similarity, concentrated mainly in the planner.
-
Discovery of novel MEV variants. Five new variants (SBA-HFT, JIT-HFT, BUR-HFT, LR-HFT, MLL-HFT) were generated and validated; the paper also notes that EVAge independently produced a mathematical proof of SBA-HFT's superiority (Appendix C).
Main Findings
-
Reference strategies reproduced: Six reference strategies spanning sandwich, backrun, and liquidity categories were reproduced and validated — the basic HFT, SBA, and JIT, and the complex BUR, LR, and MLL. Reproduction was accepted only when the planner and execution logic completed under the validator with a positive-profit case, and outputs were further reviewed by human experts.
-
Novel variants and profit uplift: Five novel variants were discovered on Ethereum, yielding a profit increase of 1.02× to 15.97× relative to their corresponding baseline strategies. Two patterns characterized the variants: larger parameter search spaces (replacing fixed or restricted parameter choices with broader searches over attacker input or liquidity size) and HFT composition (fusing HFT's search-based sizing with another strategy's action sequence).
-
Case study (SBA vs. SBA-HFT): On the same Uniswap V3 opportunity at block 21764567, a victim swapped 615.13 WETH for USDC, routing 369.02 WETH through the 0.05% USDC/WETH pool and moving the active tick from 197704 to 197817. Standard SBA used a 238,026 USDC backrun input for a net profit of 9.80 WETH (0.004 WETH gas), stopping at tick 197787. SBA-HFT used a 1,904,211 USDC input, recovering 733.36 WETH; valued at 3.477×10⁻⁴ WETH/USDC and after 0.010 WETH gas, net profit was 71.32 WETH — 7.27× the net profit of SBA. The gap is attributed to SBA's fixed multiplier list (0.5x, 1x, 1.5x) versus SBA-HFT's SizeOptimization procedure of coarse probing, ternary search, and granular local scan, which reached down to tick 197600.
-
Cross-protocol adaptation: 11 strategies were successfully adapted from CPMM to both Uniswap V3 (a CLMM) and Balancer V2 (weighted pools). EVAge autonomously introduced tick-coverage frontiers when full tick-map snapshots could not be reconstructed, and adapted shared weighted-math modules for Balancer V2's vault-mediated routing.
-
Cross-chain adaptation: Strategies were ported from Ethereum to Base and BSC. Re-adapting additional strategies to an already-supported chain incurred only marginal additional cost because agents tended to reuse the existing codebase.
-
Cost: All of the above was achieved with less than 60 dollars in LLM token costs.
-
Evaluation coverage and cost of the harness: Evaluation covers 11 MEV strategies on Ethereum, six on BSC, and two on Base, with 33 strategy-protocol lanes on Ethereum alone. The implementation comprises 5,367 lines of Python for six MEV bot workflows and 518 lines for the agent harness.
-
Assumption caveat: The authors replay each historical transaction independently, assume unbounded attacker capital, exclude private-relay submission and competition, and do not simulate asset acquisition, on-chain conversion, flash-loan repayment, or builder bribes. Reported revenues are therefore upper bounds, denominated in WETH/WBNB net of gas.
-
Planning latency and ablation results (RQ4, RQ5): The paper poses these research questions but the results are not reported in the available content; Table 2, which reports plan/candidate, simulation/plan, and profit/simulation and profit/candidate ratios, is cut off.
Methodology in Plain English
The team split an MEV bot into three jobs. A collector reads on-chain data and filters out candidate opportunities. A planner does the AMM math and decides trade sizes, parameters, and calldata. A validator actually runs the plan against a forked copy of a historical block using Anvil and reports whether it made money or reverted, and why.
Above this sits an orchestrator plus three specialist LLM agents, one per stage. The user picks a mode — variant explorer, protocol adaptation, or cross-chain adaptation — which determines which agent starts and which feedback paths are allowed. In protocol adaptation, the collector engineer starts by learning the target protocol; if the builder finds the candidates wrong, control goes back; if the validator finds the plan unprofitable, control goes back to the builder. The loop ends only when at least one historical case is confirmed profitable.
Two design choices carry most of the safety weight. First, agents only write code offline; the deployed bot executes deterministic code, so LLM latency never sits on the execution path. Second, agents are not allowed to write the Proof-of-Concept themselves — the PoC is derived strictly from the validator's concrete execution trace, so a hallucinated profit cannot be reported as a success. All experiments used the Codex CLI (v0.144.6) with GPT-5.4 at xhigh reasoning effort, on an Ubuntu 24.04.4 server with an Intel Core i9-14900, 62 GiB RAM, and a 1 TB NVMe volume, using Foundry/Anvil v1.5.1, Web3.py v7.14.1, SQLite v3.45.1, and Reth v1.10.2.
Why This Matters
Impact on research. The paper reframes MEV strategy engineering as a code-generation-and-repair problem with a deterministic oracle, rather than a purely economic or game-theoretic one. It also fills a stated gap: prior work such as Ladóczki et al. and Gogol et al. examined only CPMMs and explicitly omitted CLMMs, whereas EVAge targets CLMM tick mechanics and Balancer V2 vault routing directly.
Real-world applications:
- Democratizing searcher tooling: small teams or independent developers can port a strategy to a new protocol or chain without rebuilding tick math and calldata encoders by hand.
- Capturing launch "vacuum periods": the paper cites searchers adapting to zkSync within one month of its developer release and the race to integrate Arbitrum's TimeBoost auction; automated migration shortens that lag.
- Protocol and chain onboarding: new AMMs and L2s could ship machine-generated reference MEV implementations, or use the same harness defensively to find and mitigate exploitable patterns before deployment.
- Auditable bot maintenance: because PoCs come from execution traces rather than agent assertions, generated code carries a verifiable profitability record that reviewers can check.
Industry relevance. MEV is described as a multi-billion-dollar annual flow on Ethereum, and the paper argues its capture is concentrated among institutional players in tension with blockchain decentralization goals. Reporting full multi-strategy adaptation at under 60 dollars in token costs puts a concrete, low price on what was previously expert labor, which is directly relevant to searcher firms, market makers, L2 teams competing for order flow, and researchers studying MEV concentration.
Future Directions
- Closing the realism gap: the current evaluation assumes unbounded capital, no competing searchers, no private-relay submission, and no flash-loan or bribe modeling. Incorporating competition and capital costs would test whether the reported upper-bound revenues survive in live conditions.
- Extending beyond the evaluated protocol families: only CPMM, Uniswap V3-style CLMM, and Balancer V2 weighted pools were covered across 11 strategies. Whether the harness generalizes to order-book DEXes, intent-based systems, or further AMM designs is untested.
- Filling in the latency and ablation picture: RQ4 (planning latency across protocols and chains) and RQ5 (ablation of harness design) are posed but not reported in the available text; quantifying how much each harness component — decomposition, closed-loop routing, validator gating — actually contributes would clarify where the gains come from.
- Defensive use and disclosure norms: the paper releases five novel variants "for community reference" and includes a Proof-of-Concept workflow. How such autonomous discovery pipelines should interact with responsible-disclosure practices and with protocols' own mitigation efforts remains an open governance question.
Target Audience
Advanced readers: MEV searchers and quantitative DeFi engineers, blockchain security researchers, AMM and L2 protocol designers, and LLM-agent researchers interested in architectures that pair nondeterministic code generation with deterministic verification oracles. Readers without background in AMM invariants, tick mechanics, and bundle/sequencer semantics will find the protocol-adaptation and case-study sections demanding, though the harness design and safety argument are accessible at an intermediate level.
Authors’ abstract
Maximal Extractable Value (MEV) has evolved into a major economic force in blockchain ecosystems, yet its capture is dominated by experienced teams, and both strategy design and implementation rely on manual expert work that scales poorly across heterogeneous protocols and chains. We present EVAGE, the first fully autonomous multi-agent framework for end-to-end MEV strategy generation and adaptation. Equipped with three specialized operation modes, it automatically discovers…
The full abstract on arXiv