AgenticSTS doesn't carry an ever-growing log. It rebuilds each decision from five organized memory layers. | Image: Cheng et al.
The researchers picked the deck-building roguelike Slay the Spire 2 as their test bed. A single playthrough involves hundreds of decisions, from picking cards and planning fights to choosing routes on the map and buying items. The rules translate fully into text, randomness is high, and runs are long. Human players win 16 percent of the time on the lowest difficulty level, A0, according to the developers. Frontier models used in the AGI-Eval assessment didn't win a single game across five tested setups. The game is hard but open-ended enough that architectural differences show up clearly.
A battle in Slay the Spire 2. The agent derives its decisions from game states like this one, with cards, enemies, and stats all represented as text. | Image: Cheng et al.
Five slots replace a growing log
Typical LLM agents like ReAct or Reflexion append past observations, tool calls, and self-reflections to the next prompt. The context grows with every step until the window overflows or the model's attention gets diluted. AgenticSTS does the opposite. For each decision, the prompt is freshly built from five clearly separated slots.
