Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MACI: Multi-Agent Collaborative Intelligence for Adaptive Reasoning and Temporal Planning

Edward Y. Chang

TL;DR

MACI tackles core deficiencies of LLM-based planning by introducing a three-component framework: a Meta-Planner that constructs task-specific workflow graphs, a distributed Agent Repository of lightweight, specialized agents, and a Run-Time Monitor that adapts plans to changes. By decoupling planning from validation and enforcing restricted context windows, MACI mitigates metacognitive shortcomings, attention bias, and common-sense gaps, enabling System-2 level reasoning in temporal planning. The approach is validated on Traveling Salesperson Problem and Thanksgiving Dinner Planning, showing improved constraint satisfaction, adaptability, and planning robustness versus independent LLMs. The work demonstrates that structured meta-planning, distributed validation, and proactive multi-agent coordination can significantly improve real-world planning under uncertainty, providing a scalable path toward more reliable AI-assisted decision-making.

Abstract

Artificial intelligence requires deliberate reasoning, temporal awareness, and effective constraint management, capabilities traditional LLMs often lack due to their reliance on pattern matching, limited self-verification, and inconsistent constraint handling. We introduce Multi-Agent Collaborative Intelligence (MACI), a framework comprising three key components: 1) a meta-planner (MP) that identifies, formulates, and refines all roles and constraints of a task (e.g., wedding planning) while generating a dependency graph, with common-sense augmentation to ensure realistic and practical constraints; 2) a collection of agents to facilitate planning and address task-specific requirements; and 3) a run-time monitor that manages plan adjustments as needed. By decoupling planning from validation, maintaining minimal agent context, and integrating common-sense reasoning, MACI overcomes the aforementioned limitations and demonstrates robust performance in two scheduling problems.

MACI: Multi-Agent Collaborative Intelligence for Adaptive Reasoning and Temporal Planning

TL;DR

MACI tackles core deficiencies of LLM-based planning by introducing a three-component framework: a Meta-Planner that constructs task-specific workflow graphs, a distributed Agent Repository of lightweight, specialized agents, and a Run-Time Monitor that adapts plans to changes. By decoupling planning from validation and enforcing restricted context windows, MACI mitigates metacognitive shortcomings, attention bias, and common-sense gaps, enabling System-2 level reasoning in temporal planning. The approach is validated on Traveling Salesperson Problem and Thanksgiving Dinner Planning, showing improved constraint satisfaction, adaptability, and planning robustness versus independent LLMs. The work demonstrates that structured meta-planning, distributed validation, and proactive multi-agent coordination can significantly improve real-world planning under uncertainty, providing a scalable path toward more reliable AI-assisted decision-making.

Abstract

Artificial intelligence requires deliberate reasoning, temporal awareness, and effective constraint management, capabilities traditional LLMs often lack due to their reliance on pattern matching, limited self-verification, and inconsistent constraint handling. We introduce Multi-Agent Collaborative Intelligence (MACI), a framework comprising three key components: 1) a meta-planner (MP) that identifies, formulates, and refines all roles and constraints of a task (e.g., wedding planning) while generating a dependency graph, with common-sense augmentation to ensure realistic and practical constraints; 2) a collection of agents to facilitate planning and address task-specific requirements; and 3) a run-time monitor that manages plan adjustments as needed. By decoupling planning from validation, maintaining minimal agent context, and integrating common-sense reasoning, MACI overcomes the aforementioned limitations and demonstrates robust performance in two scheduling problems.

Paper Structure

This paper contains 119 sections, 22 equations, 19 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Limitations of LLMs in Planning
  3. The MACI Framework
  4. Summary: How MACI Addresses LLM Limitations
  5. Related Work
  6. Case Study: Illuminating LLM Limitations
  7. Problem Statement
  8. Study #1 Sequential Planning
  9. Diagnoses: Common-Sense Constraints and Isolated Processing Syndrome
  10. Study #2 Reactive Planning
  11. Diagnosis: Attention Narrowing
  12. Summary of LLM Limitations in Planning
  13. Metacognitive Limitations
  14. Attention Bias
  15. Common Sense Gaps
  16. ...and 104 more sections