Large Reasoning Models (LRMs) employing Chain-of-Thought (CoT) reasoning have shown remarkable capabilities in solving various problems, but they frequently suffer from degenerate hallucination loops when confronted with complex tasks. In standard continuous generation setups, these models often produce excessively long, messy, and logically meaningless thinking blocks that exhaust context limits without reaching a valid conclusion. While existing methods attempt to mitigate this by filtering the final output or relying on outcome-based verification, they fail to structurally intervene within the flawed intermediate logic. To mitigate these degenerate reasoning loops and context exhaustion, we propose TRACER: Trace Refinement and Condensation for Enhanced Reasoning. Rather than allowing a model to generate thoughts and responses in a single unmanaged stream, the TRACER architecture injects a dedicated editing agent into the reasoning process. This approach partitions the autoregressive process into three distinct phases: drafting the exploratory logic (Thinker), auditing the sequence for errors (Editor), and formulating the final output (Responder).

We evaluate TRACER across rigorous mathematical (AIME, MATH500) and logical (MMLU, ZebraLogic) benchmarks and demonstrate that injecting an Editor can reduce the hallucination behavior. Our empirical results reveal that this architecture drastically condenses reasoning traces and achieves substantial absolute accuracy improvements, especially on complex tasks such as ZebraLogic and AIME. Furthermore, we establish that architectural homogeneity between the drafting and auditing agents yields superior weak-to-strong generalization compared to computationally stronger, heterogeneous models. Ultimately, this work proves that actively managing test-time compute through structured, multi-agent trace condensation is an effective framework for ensuring robust intrinsic reasoning.