LLM-powered agents running in production skip self-verification nearly 98% of the time, according to a new behavioral analysis of hundreds of real execution traces.
A team of researchers encoded the runtime behavior of a live ReAct agent system into four-letter symbolic sequences - X for Explore, E for Execute, P for Plan, V for Verify - then applied pattern-mining techniques borrowed from genomics to 347 traces collected over eight days. The standout finding: agents transition from execution to verification only 2.1% of the time, suggesting a systemic habit of acting without checking results. A three-step loop pattern - Plan, Explore, Plan again - was the only statistically significant risk signal, correlating with a 10.4% drop in task success rates. The researchers validated these patterns against 2,000 public SWE-agent traces, finding the same deficits replicated in an entirely independent system.
Most efforts to improve agent reliability focus on better models or prompts; this paper argues the problem can be caught and corrected at runtime, without touching the underlying model at all. The team's "Governor" system - a layered combination of a rule engine, statistical accumulator, and chi-square threshold adaptor - delivered a 6.2 percentage-point improvement in task success and cut average token consumption by 44% in a production before-and-after test. For teams already paying per-token at scale, that reduction is the most immediately legible argument for adding this kind of monitoring.
The genomics analogy is mostly decorative - n-gram analysis and Markov chains do not require the DNA metaphor to be useful. But the underlying empirical finding is harder to dismiss: agents, left unmonitored, spin in planning loops and skip their own sanity checks at a rate that would concern anyone running them in production.