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

Researchers have proposed Base Sequence Analysis, a framework that encodes the runtime behavior of large language model (LLM) agents into symbolic sequences to identify failure patterns, and a companion intervention system called Governor that improved task success while cutting token use, according to a paper posted to arXiv [1]. The framework uses a four-letter alphabet — X for Explore, E for Execute, P for Plan, and V for Verify — to represent agent actions, drawing an analogy to genomic sequence analysis [1]. The authors applied n-gram pattern mining, Markov transition matrices, and point-biserial correlation to 347 real-world execution traces collected from a production ReAct agent system over 8 days [1]. The analysis found that the trigram P-X-P was the only statistically significant high-risk pattern, lowering the success rate by 10.4% [1]. The P-ratio emerged as the strongest negative predictor of success, with a correlation of r=-0.256 and p<0.0001 [1]. The transition probability from Execute to Verify was just 2.1%, indicating what the paper describes as a systemic verification deficit [1]. Based on these findings, the team designed Governor, a three-layer runtime intervention system consisting of a rule engine, a statistical accumulator, and a chi-square-based threshold adaptor [1]. In a natural before-and-after deployment evaluation spanning 101 tasks before and 246 tasks after deployment, Governor delivered a 6.2% absolute increase in task success rate while reducing average token consumption by 44% [1]. To test whether the behavioral patterns generalized beyond the original system, the researchers applied the XEPV encoding to 2,000 public SWE-agent trajectories on the SWE-bench benchmark [1]. The analysis confirmed that exploration spirals and the E-to-V verification deficit replicated in an independent agent system [1]. The work arrives as LLM-powered autonomous agents are being integrated into software development and research workflows. Google's Gemini models, for instance, have introduced agentic capabilities for autonomous coding and research tasks, with later generations focused on reducing hallucinations and improving latency [4]. The paper outlines six future research directions, including base sequence language models, cross-agent behavioral fingerprinting, and reward shaping, and the authors have released an open-source toolkit for reproducibility [1].