MetaCogAgent: A Metacognitive Multi-Agent LLM Framework with Self-Aware Task Delegation

2026-05-19 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

MetaCogAgent is a multi-agent large language model (LLM) framework that integrates metacognitive capabilities, enabling agents to assess their own competence and adaptively delegate tasks. Inspired by cognitive science, each agent features a Metacognitive Self-Assessment Unit that combines verbalized uncertainty with historical capability profiles to estimate per-task confidence. When confidence is low, an adaptive delegation protocol routes tasks to more suitable agents through cross-agent evaluation. A capability boundary learning module refines each agent's competence model iteratively via cybernetic feedback. Evaluated on the MetaCog-Eval benchmark, which comprises 700 tasks across five cognitive dimensions, MetaCogAgent achieved 82.4% task accuracy, an 8.7% improvement over the best routing baseline. It also demonstrated efficiency, using 5% fewer API calls than AutoGen and 34% fewer than ensemble voting, with an Expected Calibration Error (ECE) of 0.087.

Key takeaway

For AI Engineers building multi-agent LLM systems, integrating metacognitive self-assessment and adaptive delegation can significantly boost performance and efficiency. You should consider implementing a confidence-gated delegation mechanism, especially for complex or cross-domain tasks, to ensure tasks are handled by the most competent agent. This approach reduces API calls and improves accuracy, particularly on challenging problems, by preventing agents from overconfidently executing tasks beyond their expertise.

Key insights

Metacognitive multi-agent LLM systems improve task accuracy and efficiency by enabling self-assessment and adaptive delegation.

Principles

• Combine verbalized uncertainty with historical profiles for calibrated confidence.
• Delegate tasks when self-assessed confidence falls below a threshold.
• Refine agent competence models via cybernetic feedback loops.

Method

Agents self-assess task confidence, delegate low-confidence tasks to peers, and update capability profiles based on performance feedback, forming a cybernetic loop.

In practice

• Implement self-assessment units in multi-agent LLM systems.
• Use adaptive delegation for complex or cross-domain tasks.
• Employ feedback loops to refine agent competence over time.

Topics

• Multi-Agent LLM Systems
• Metacognition
• Self-Aware Task Delegation
• Capability Boundary Learning
• MetaCog-Eval Benchmark

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• MIRROR: A Hierarchical Benchmark for Metacognitive Calibration in Large Language Models — LLMs possess partial self-knowledge but universally fail to translate it into appropriate agentic action without external architectural control.
• Week Ending 3.22.2026 — Advancements in LLM reasoning, multi-agent coordination, creativity, and social choice highlight diverse progress in AI capabilities and understanding.
• CAPRA: Scaling Feedback on Software Architecture Deliverables with a Multi-Agent LLM System — Multi-agent LLM systems can automate complex software architecture feedback by combining specialized agents and robust verification.
• Wisdom Of The (AI) Crowd: Investigating Artificial Swarm Intelligence In Large Language Models — LLM swarms can reduce estimation errors and exhibit metacognitive awareness, mimicking human swarm intelligence.
• DecisionBench: A Benchmark for Emergent Delegation in Long-Horizon Agentic Workflows — Emergent delegation in LLM agents is best evaluated via process-level metrics, as end-task quality alone obscures orchestration signals.
• Connect the Dots: Training LLMs for Long-Lifecycle Agents with Cross-Domain Generalization Via Reinforcement Learning — LLMs can achieve continuous self-improvement and cross-domain generalization by learning to self-update context via end-to-end RL.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.