When Agents Disagree With Themselves: Measuring Behavioral Consistency in LLM-Based Agents
Aman Mehta
TL;DR
The paper investigates how consistently LLM-based agents behave when given identical tasks, revealing substantial divergence in action sequences across repeated runs. Using a ReAct-style agent across three models and 3,000 trials on HotpotQA, it shows that 2.0–4.2 unique action sequences per 10 runs are common, and that higher behavioral consistency strongly predicts correctness (80–92% accuracy for consistent tasks vs 25–60% for inconsistent tasks). It identifies step 2 (the first search query) as the primary divergence point and demonstrates that shorter, more consistent trajectories correlate with better outcomes, while longer paths indicate uncertainty and lower accuracy (r = -0.34). Temperature reduction improves both consistency and accuracy, suggesting a practical lever for deployment. The findings motivate runtime consistency monitoring as a diagnostic tool for early error detection and reliability improvements in complex, tool-using agents, and point to design and evaluation strategies that account for behavioral variance in real-world applications.
Abstract
Run the same LLM agent on the same task twice: do you get the same behavior? We find the answer is often no. In a study of 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, and Claude Sonnet 4.5) on HotpotQA, we observe that ReAct-style agents produce 2.0--4.2 distinct action sequences per 10 runs on average, even with identical inputs. More importantly, this variance predicts failure: tasks with consistent behavior ($\leq$2 unique paths) achieve 80--92% accuracy, while highly inconsistent tasks ($\geq$6 unique paths) achieve only 25--60%, a 32--55 percentage point gap depending on model. We trace variance to early decisions: 69% of divergence occurs at step 2, the first search query. Our results suggest that monitoring behavioral consistency during execution could enable early error detection and improve agent reliability.