Your AI agent nailed the financial reconciliation today. Will it do the same thing tomorrow? Probably not, and that is a massive problem for production environments. New data reveals a troubling reliability blind spot in how we evaluate large language model agents. While standard benchmarks report average success rates, they mask a significant volatility issue. On the AppWorld benchmark, a ReAct agent powered by GPT-4.1 achieved a respectable 77.4% mean success rate across five runs. However, it only succeeded on all five attempts for 53.0% of tasks. This 24.4-point difference is what researchers call the “consistency gap,” and it represents tasks where the model’s performance is essentially a coin flip.
This instability stems from the shape of the model’s probability distributions at critical decision points. When an agent chooses an API call or argument, it selects from a range of possible next tokens. If the distribution is “sharp,” one option dominates, leading to reliable outputs. But if the distribution is “flat,” several options have nearly equal probability. Minor computational noise—like GPU floating-point variations or request batching—can tip the scale toward a different token. Since agent workflows chain dozens of these decisions, small per-step variances compound into major trajectory shifts. This happens even at temperature zero, proving that deterministic decoding settings cannot fully eliminate randomness in hosted endpoints.
To address this, the team behind ALTK-Evolve introduced the Consistency Analyzer. Unlike traditional methods that require ground truth or extensive re-running, this diagnostic tool resamples specific decision points within a single recorded trajectory. It identifies “flip-prone” steps where the model was one token away from making a different choice. By converting these diagnostics into “consistency guidelines” and injecting them back into the inference process, the system halves the consistency gap from 24.4 points to 12.0 points. Crucially, this improvement comes without sacrificing average accuracy. For enterprises deploying agents in high-stakes roles, shifting focus from Mean@k to Pass^k metrics is no longer optional—it is essential for building trust.
