Your Agent Has a Genome: Sequence-Level Behavioral Analysis and Runtime Governance of LLM-Powered Autonomous Agents

In three linesBase Sequence Analysis framework encodes LLM-powered autonomous agent behavior into symbolic sequences (X/E/P/V). Analysis of 347 production ReAct traces reveals P-X-P pattern reduces success by 10.4% and P-ratio negatively predicts success (r=-0.256). Governor runtime intervention system achieves +6.2% absolute success increase and 44% token reduction. Validated on 2,000 SWE-agent trajectories.

Summary generated by Claude — human-verified