Most agentic reinforcement learning research assumes single-session trajectories, where an agent interacts within one continuously accumulated context. A new arXiv paper, "Mosaic: Multi-Session Agentic Reinforcement Learning," argues that this assumption is limiting and introduces a method called Mosaic for multi-session settings.
The abstract cites agentic techniques such as context compaction and subagents as motivations for moving beyond a single session. However, the provided abstract is truncated mid-sentence, so the full details of the Mosaic method—how it handles session boundaries, memory, or training—are not available in this source.
The significance of the work lies in addressing a realistic gap: many deployed agents operate across separate sessions rather than one endless context. If the paper delivers on its premise, it could push agentic RL toward more practical, long-horizon scenarios. But without the full text, the specifics remain unclear.