LLMs are increasingly deployed as orchestrators that coordinate specialized subagents to solve complex tasks. In many important domains (game playing, robotics, autonomous driving), the strongest available agents are not language models: their expertise is encoded in latent representations of policy, value, and learned features, not in words. Integrating such agents into LLM reasoning pipelines currently requires verbalization: compressing rich continuous representations into sparse textual summaries at each interaction step. To study whether verbalization constitutes a structural bottleneck, we introduce LLAMIA-Bench, a suite of six collaborative chess tasks spanning three facets: behavioral imitation, state assessment, and natural-language explanation. Tasks are drawn from prominent problems in chess literature and industry, each unsolvable by either component alone, and selected to vary in how much relevant signal survives verbalization.

Our experiments reveal a consistent verbalization debt: the performance gap between verbalized and internalized integration widens throughout training and persists as the LLM scales from 4B to 14B parameters. The effect is sharpest on tasks whose target signal has no faithful text representation: on puzzle interest prediction, verbalized systems of all sizes remain near-chance (≤ 12/100) while internalization reaches 52/100. To close this gap, we introduce latent state internalization, which projects the subagent’s continuous representations directly into the LLM’s token stream as learned state tokens, with dynamic re-encoding as actions advance the environment state. A single 14B model, LLAMIA, trained under this paradigm matches or exceeds dedicated task specialists and frontier models including GPT-5.1 with tool access across all benchmark tasks, and generalizes out-of-distribution where task-specific finetunes collapse. In a human study with skilled chess players, LLAMIA passes as human in 61% of gameplay trials and its commentary is preferred in 72.2% of pairings.