Thought Collapse in LLM Agents: A Cross-Task Diagnosis of Capacity Domination in Agentic Reasoning

The ReAct framework prepends an explicit Thought step to every agent action, and most agent frameworks treat that Thought as load-bearing. We test the assumption with a three-layer diagnosis on five models (Qwen2.5-14B/32B, DeepSeek-V3, Llama-3.1-70B, and a Llama-3.1-8B boundary cell) and five tasks spanning low to high parametric coverage (GSM8K, MATH-hard, HotpotQA, WebQuestions, TriviaQA). The central finding is Capacity Domination: when a model's capacity exceeds a task's coverage in pre-training, the Thought step collapses from a useful reasoning step into a redundant or actively harmful one. Behaviorally, the Thought-Gap (the EM gap between Full and No-Thought) decays from +0.59 on low-coverage GSM8K (14B) to non-significant for 32B on high-coverage WebQuestions and TriviaQA. DeepSeek-V3 retains a small residual gap, consistent with a two-stage collapse (scaffold degradation precedes Thought-level collapse), and cross-family Llama-3.1-70B shows the same decay, declining monotonically across all five coverage levels into negative-Gap territory. A Direct Probe that strips the entire ReAct scaffold outperforms full ReAct by up to 4× on high-coverage tasks. A Thought Causal Influence (TCI) probe records action–Thought Mismatch of 0.62–0.84 on high-coverage tasks with Follows-Q near 100%: the Action conditions on the question, not on the Thought. Motivated by the diagnosis, we propose Coverage-Aware Routing (CAR), which estimates coverage online and routes between Direct and ReAct, recovering up to +0.225 EM while saving 94–99% of Thought tokens. Our analysis is scoped to single-tool, short-horizon agents; whether the same collapse persists under multi-tool, long-horizon composition is an open question.