Bounded LLM Orchestration With Deterministic Safety Gates for Concentrated Liquidity Management

Concentrated liquidity management (CLM) in automated market makers such as Uniswap v3 requires continuous decision-making under price volatility, gas-cost uncertainty, and partial observability. Prior work has shown that reinforcement learning (RL) agents can outperform heuristic strategies under realistic cost assumptions. However, deploying end-to-end RL agents in decentralized production systems remains difficult because practical CLM automation must also satisfy interpretability, governance, and operational-safety requirements. This paper presents an LLM-orchestrated agent architecture for CLM in which a large language model (LLM) serves strictly as a decision-layer orchestrator. The model reasons over structured observations of vault and market state to recommend bounded liquidity actions, while custody and execution remain deterministically enforced by on-chain smart contracts. We evaluate the system using two evidence layers: (i) a controlled replay over the full year of 2025 under matched execution semantics, and (ii) ongoing live-deployment follow-up covering 351 production decisions across two Base Mainnet ETH/USDC vaults over 24 days (2026-01-18 to 2026-02-11). The replay layer provides broader regime coverage, while the live layer validates operational safety behavior in production. Results show conservative action selection, strong boundary responsiveness, and deterministic fail-closed behavior through schema validation and execution gating.