Base Sequence Analysis and Governor: +6.2% agent success

Sidi Deng submitted a paper on 5 Apr 2026 proposing Base Sequence Analysis, a compact sequence encoding and runtime governance approach for LLM-powered autonomous agents. The method encodes agent runtime behavior into a four-letter XEPV alphabet (X Explore, E Execute, P Plan, V Verify) and applies sequence mining to 347 real-world execution traces collected over 8 days from a production ReAct agent system.

What did the study analyze and find?

The paper analyzed 347 execution traces from a production ReAct agent system over 8 days and identified specific sequence-level risks: the trigram P-X-P is the only statistically significant high-risk pattern, lowering task success rate by 10.4%, P-ratio correlates negatively with success (r = -0.256, p < 0.0001), and the E->V transition probability is only 2.1%, indicating a systemic verification deficit. The authors used n-gram pattern mining, Markov transition matrices, and point-biserial correlation on the XEPV-encoded traces to arrive at these numbers.

The paper also applied the same XEPV encoding to 2,000 public SWE-agent trajectories on SWE-bench and found that exploration spirals and the E->V verification deficit replicate in an independent system, supporting cross-system generality.

How does Governor work and what effect did it have?

Governor is a three-layer runtime intervention system composed of a rule engine, a statistical accumulator, and a chi-square-based threshold adaptor, designed to intervene based on sequence-level signals. The rule engine enforces explicit checks, the statistical accumulator gathers runtime statistics across traces, and the chi-square-based threshold adaptor sets thresholds for interventions.

The authors evaluated Governor in a natural before/after deployment with N = 101 before and N = 246 after. Deployment of Governor produced a +6.2% absolute increase in task success rate and simultaneously reduced average token consumption by 44%. These changes were measured on the same production ReAct agent system used for the sequence analysis.

Why it matters

Sequence-level behavior provides a compact, interpretable signal that links agent decision phases (plan, explore, execute, verify) to measurable outcomes. Identifying a single high-risk trigram that cuts success by 10.4% offers a clear target for runtime checks, and the 2.1% E->V transition probability exposes a practical verification gap that governance systems can address. Governor’s reported +6.2% success boost alongside a 44% token reduction suggests interventions can improve both effectiveness and efficiency at runtime, not just in offline evaluation.

These findings matter for teams building LLM-powered autonomous agents because they move governance from coarse metrics into sequence-aware, real-time controls that map directly to agent behaviors observed in production.

What to watch

The paper outlines six follow-up research directions, including base sequence language models, cross-agent behavioral fingerprinting, and reward shaping; adoption of any of these will test whether sequence-level methods generalize beyond the studied systems. A near-term signal will be community uptake of the authors’ open-source toolkit and replication of Governor-style interventions on other public benchmarks such as SWE-bench.

Summary of concrete data points from the paper: the study used 347 real-world execution traces collected over 8 days; the P-X-P trigram lowered success by 10.4%; P-ratio correlated with success at r = -0.256 (p < 0.0001); the E->V transition probability was 2.1%; Governor’s before/after evaluation used N = 101 vs N = 246 and produced a +6.2% absolute increase in task success and a 44% reduction in average token consumption.

References: Sidi Deng, "Your Agent Has a Genome: Sequence-Level Behavioral Analysis and Runtime Governance of LLM-Powered Autonomous Agents," arXiv:2606.15579, submitted 5 Apr 2026. The paper includes an open-source toolkit for reproducibility.