AgentStream: How Well Do Self-Evolving LLM Agents Perform Under Streaming Tasks?

AgentStream introduces a unified framework for evaluating self-evolving LLM agents on configurable task streams. It defines Isolated, Sequential, and Interleaved test scenarios that progressively mix task scope and domains. The study combinatorially evaluates five representative self-evolving methods across three frontier foundation models. According to the provided abstract, self-evolution reliability depends strongly on the streaming scenario, its gains are gated by model capability and vary non-monotonically with model strength, and no single method consistently dominates across models and scenarios.