A new lightweight parallel monitoring architecture known as the Cognitive Companion detects and recovers from reasoning degradation such as looping, drift and stuck states in LLM agents.

The Cognitive Companion runs alongside primary agent execution with two variants: an LLM-based monitor that reduced repetition on loop-prone tasks by 52-62% at 11% overhead and a probe-based monitor trained on hidden states from Gemma 4 layer 28 that delivered mean effect size +0.471, cross-validated AUROC 0.840 and zero measured inference overhead (https://arxiv.org/abs/2604.13759). This addresses error accumulation rates up to 30% on hard multi-step tasks. ReAct introduced interleaved reasoning and acting yet left runtime degradation unmonitored (https://arxiv.org/abs/2210.03629); the original paper's feasibility study identifies task-type sensitivity—strongest gains on open-ended and loop-prone tasks, neutral or negative on structured ones—as a constraint few capability papers address.

AgentBench evaluations previously quantified high failure rates in web and tool-use agents but focused on benchmark design rather than internal state probes for live recovery (https://arxiv.org/abs/2308.08998). The probe approach synthesizes Reflexion-style self-correction (https://arxiv.org/abs/2303.11366) with sub-token monitoring to avoid per-step LLM judges that impose 10-15% overhead. Small-model tests on Qwen 2.5 1.5B and Llama 3.2 1B found no quality gains even when probes fired, revealing a probable scale boundary below 2B parameters missed by most agentic-AI scaling literature.

Selective activation keyed to task type and scale emerges as the practical path forward; current research emphasis on prompting and larger models has largely overlooked architectural safeguards against the core fragility now measurable in production agent deployments.