AEP v1.1 Claims 80% Token Reduction by Adding Microkernel Runtime to LLM Agents

A proposed architecture called Agent Execution Protocol (AEP) v1.1 aims to fix structural reliability problems in current LLM agent frameworks, which store operational state inside the chat context rather than a dedicated runtime. The approach introduces an 8-register microkernel sandbox — built in Python, Rust, or Go — that manages watchdog timers, transaction rollback logs, and context budgets entirely outside the LLM's awareness. Key resilience features include loop detection via sandbox state hashing, ACID-style rollbacks using a write-ahead log, and structured error injection so the model can self-correct without corrupting the filesystem. In a controlled benchmark using Claude 3.5 Sonnet on a 20-spreadsheet transformation task (n=50 per arm), the AEP runtime reportedly cut token consumption by nearly 80% and wall time by 74%, while eliminating execution loops and file corruption entirely. The authors caution that results cover only a spreadsheet pipeline so far, and have not yet been validated on code generation, web scraping, or other agent workloads.