Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital

Operating-Layer Controls for Onchain LLM Agents Under Real Capital Researchers deployed 3,505 LLM-powered trading agents over 21 days on Ethereum, executing 300K onchain transactions across roughly $20M volume with 99.9% settlement success on policy-valid transactions. The system generated 70B inference tokens and 6,000+ sequential prompt-state-action cycles per long-running agent, stress-testing reliability at production scale. Reliability was achieved through structured operating-layer controls layered above the base model, not model capability alone. Autonomous agent developers and DeFi protocol engineers should study this architecture closely as the first large-scale empirical trace linking user mandates to validated onchain settlement under real financial risk.