COSMIC: Leader-Driven Context-Oriented Collaboration Between Agents

Vansh Vashishtha

doi:10.1109/TCSS.2026.3661839

COSMIC: Leader-Driven Context-Oriented Collaboration Between Agents

2026-02-27

0

PRE

AI

DOI：10.1109/TCSS.2026.3661839

Original

Original request for help

Save

Abstract

En 中文

We present COSMIC, a leader-driven multiagent coordination framework that extends the AutoGen GroupChat paradigm to support more stable and context-consistent collaboration among large language model (LLM) agents. COSMIC introduces: 1) a leader agent that decomposes tasks and issues an explicit termination decision; 2) a shared rolling memory buffer that synchronizes recent context across agents; and 3) a speaker selector agent (SSA) that enforces structured turn-taking and mitigates self-invocation and speaker imbalance. Across six coordination-intensive tasks, COSMIC achieves 100% validity under task-specific hard-constraint criteria. On the five tasks where baseline frameworks produced validated outputs, COSMIC improves composite scores by 20.2%, 23.3%, 27.7%, 52.6%, and 46.9% over CrewAI Hierarchical, CrewAI Sequential, AutoGen, LangChain, and OpenAI Swarm, respectively. COSMIC reduces wall-clock time relative to CrewAI Hierarchical, CrewAI Sequential, AutoGen, and LangChain by 57.3%, 50.8%, 61.4%, and 77.6%, while incurring higher runtime than OpenAI Swarm due to richer coordination and memory overhead. On GSM8K, COSMIC attains $94\pm 1$% accuracy, exceeding the sparse multiagent debate (MAD) configuration reported in a Google Research publication. Empirically, performance peaks around five agents, reflecting a trade-off between diversity and coordination overhead.

Keywords:

Agent communication

AutoGen

collaborative AI

leader-driven termination

multiagent coordination

shared memory

COSMIC: Leader-Driven Context-Oriented Collaboration Between Agents

Abstract

Journal

Organization

Cited Papers