AI 代理與評測
Self-Improving Agents Face Random Task Orders, Reversing the Average Gain From +1.5% to −4.5%
Salesforce reevaluated two text-based memory methods and found that adding a self-improvement loop increased run-to-run variance in 71% of configurations. Once the fixed task order—which implicitly formed an easy-to-hard curriculum—was shuffled, ReasoningBank not only failed to improve the agent but reduced its success rate.
Salesforce AI Research reevaluated Agent Workflow Memory (AWM) and ReasoningBank, two methods in which an agent writes successful workflows or reasoning experiences to a text-based memory store after completing each task, allowing them to be retrieved for subsequent tasks. The team used GPT-5-mini as both the agent and the memory-construction model, testing it on WebArena, VisualWebArena, and the enterprise CRM environment SCUBA. It also expanded the commonly used single run with a fixed task order to three runs per configuration and two randomized orderings.
The results show that memoryless agents already exhibit non-negligible randomness. After adding a stateful memory loop, differences in early samples can be amplified along the task sequence. Across the 24 domain-method configurations compared, the standard deviation increased in 17—about 71%—with 11 showing increases of more than 50%. On WebArena’s GitLab subset, the gap between the best and worst runs widened from 4.44 percentage points for the baseline to 7.78 percentage points with ReasoningBank; in some domains, the gap exceeded 10 points. This means that running an evaluation only once and reporting an improvement of a few percentage points could easily mistake run noise for a genuine method-level gain.
More consequential was the effect of task order. The default ordering implicitly formed an easy-to-hard curriculum, under which ReasoningBank produced an average gain of 1.5%. After randomization, however, it caused an average decline of 4.5%. Manual inspection found that the agent could turn underspecified experiences into rules that sounded plausible but were not executable—for example, recommending direct API calls in an environment where it could interact only through a browser. Adding detailed scoring criteria, environment feedback, and capability constraints to memory construction recovered about 31% of the order-related degradation, but did not eliminate the gap.
From an engineering perspective, long-term memory should therefore be treated as mutable state that can drift, rather than as a free capability upgrade. Evaluations should report results from multiple runs and randomized task orders, and should audit memory contents. Production systems should also provide interfaces for deleting or correcting erroneous experiences. The authors have released the experimental code and 39,343 agent trajectories. However, the findings are currently limited to three web environments, three repeated runs, and settings that use real task rewards; they cannot yet be directly generalized to coding agents or production systems without reliable feedback.