LLM agents that condition on live user state often suffer from direction drift: responses are grammatically correct but misaligned with current state. The authors propose State-Grounded Conditioning (SGC), a design principle that externalizes state-dependent control into three wrappers (Perception, Grounding, Interaction) that operate on structured state slices.
Evaluated on a 200-session benchmark from an in-game coaching agent, SGC raises turn-level grounded accuracy from ~61-70% to 96.7% and session-level from ~20-26% to 83.5%. It reduces mean first-token latency from 6.1s to 1.5s and cuts session-level grounding failures by about 78% compared to the strongest baseline.
SGC addresses a practical need for state-dependent LLM agents in latency-sensitive, user-facing applications. The wrappers are designed to be complementary and approximately orthogonal, providing a principled alternative to monolithic LLM handling of live state.