AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

Researchers have developed AgenticSTS, a testbed for long-horizon LLM agents that uses a bounded contract approach to improve performance in complex decision-making tasks. The testbed was evaluated in the game Slay the Spire 2 and showed improved performance with the addition of strategic skills. The researchers released a reproducible testbed with 298 completed trajectories and analysis scripts.

1sec.ai