msc team

LLM agents pay for every step they take. ATISBO-SPEC lets an agent skip calls when the next steps are predictable. A small model of my own (Atisbo, a GRU, not a Meta model) predicts the agent's next three actions. While it is confident, the agent acts without calling the LLM; when it is not, or when the world stops matching its prediction, Muse Spark decides. SAM 3.1 turns each frame into masks, and the loop only sees what it rebuilds from them. Over 50 episodes in a simulated 12 × 12 world, Muse Spark calls fell from 92 to 49 per episode, with 96% success against 100%. With SAM in the loop, success fell to 75% over 20 episodes. My hypothesis of a 70% cut failed, and the repo reports it with confidence intervals, ablations and an offline replay of every result. A live console shows each real call to Meta's API, and a public page replays recorded episodes.