MorphAgent: Empowering Agents through Self-Evolving Profiles and Decentralized Collaboration

Abstract: LLM based multi-agent systems (MAS) have shown promise in tackling complex tasks, but often rely on predefined roles and centralized coordination, limiting their adaptability to evolving challenges. This paper introduces MorphAgent, a novel Autonomous, Self-Organizing, and Self-Adaptive Multi-Agent System for decentralized agent collaboration that enables agents to dynamically evolve their roles and capabilities. Our approach employs self-evolving agent profiles, optimized through three key metrics, guiding agents in refining their individual expertise while maintaining complementary team dynamics. MorphAgent implements a two-phase process: a Profile Update phase for profile optimization, followed by a Task Execution phase where agents continuously adapt their roles based on task feedback. Our experimental results show that MorphAgent outperforms existing frameworks in terms of task performance and adaptability to changing requirements, paving the way for more robust and versatile multi-agent collaborative systems.

• The paper demonstrates a decentralized MAS where agents autonomously update profiles using metrics for role clarity, differentiation, and task-role alignment.
• It introduces a dynamic profile mechanism that enables self-organization and continuous adaptation, improving resilience against node failures.
• Empirical evaluations across code generation, reasoning, and mathematical tasks reveal significantly higher accuracy compared to centralized frameworks.

Decentralized Adaptation in Multi-Agent Systems: MorphAgent Overview

Motivation and Core Principles

MorphAgent (2410.15048) addresses critical limitations in LLM-based Multi-Agent Systems (MAS), particularly the lack of adaptability, constrained role formation, and reliance on centralized task coordination. Drawing inspiration from decentralized natural swarm systems, the authors distill three foundational principles for robust team performance: Individual Autonomy, Self-Organization, and Self-Adaptability. Existing MAS frameworks either constrain agent autonomy via rigid workflows, maintain organizational structures through centralized coordinators, or restrict agent adaptability to externally triggered modifications. MorphAgent advances the state-of-the-art by enabling agents to independently evolve their profiles, thus supporting dynamic collaboration and resilience in complex, changing environments.

Architecture: Dynamic Profile Mechanism and Decentralized Collaboration

MorphAgent's architecture centers on a dynamic agent profile mechanism: each agent maintains a profile encoding its functional capabilities, context-dependent adjustments, and interaction preferences. Crucially, profiles are mutable, evolving continuously according to local observations, task requirements, and peer interactions.

Formally, at time $t$, agent $a_i$ updates its profile via a function $\psi$ implemented by an LLM: $$p_i^t = \psi(o_i^t, p_i^{t-1}, \{\alpha_j^t\}_{j \in \mathcal{N}_i^t})$$ where $o_i^t$ are observations, $p_i^{t-1}$ is the previous profile, and $\{\alpha_j^t\}$ represent neighboring agents' actions. This enables agents to autonomously adjust both their roles and their participation in team strategies.

MorphAgent alternates between two phases:

• Profile Update: Agents iteratively optimize their profiles using three metrics—Role Clarity Score (RCS), Role Differentiation Score (RDS), and Task-Role Alignment Score (TRAS)—and receive context-dependent prompts for improvement or specialization.
• Task Execution: Agents act based on the optimized profiles, adapting further in response to feedback and environmental changes.

This cycle ensures both individual specialization and dynamic collaboration without a centralized controller or static workflow.

Profile Optimization Metrics

Effective decentralization and adaptability are achieved by systematically optimizing three key metrics:

• Role Clarity Score (RCS): Quantifies unambiguity and explicitness of an agent’s profile using syntactic complexity, lexical diversity, and skill relevance.
• Role Differentiation Score (RDS): Measures profile diversity, encouraging high specialization and reducing redundant agent behaviors by maximizing profile embedding dissimilarity.
• Task-Role Alignment Score (TRAS): Evaluates how closely an agent’s profile matches current task requirements, integrating semantic similarity and capability matching.

Agents refine their profiles iteratively via feedback prompts driven by changes in these metrics. Using all three metrics jointly yields the highest benchmark performance, as evidenced by ablation experiments.

Empirical Evaluation and Contradictory Claims

MorphAgent is systematically evaluated on code generation (BigCodeBench), general reasoning (BigBenchHard), and mathematical reasoning (MATH). Three-agent teams are tested under dynamic environments, including stochastic node (agent) failures. Results demonstrate:

• Superior robustness: With high failure probability (0.8), MorphAgent achieves 40.10–54.00% accuracy, whereas baselines degrade to 1.43–19.52%, a strong numerical result directly contradicting claims of resilience in SOP-based or centralized MAS.
• Strong benchmark performance: Outperforms baselines across all tasks and LLM backbones.
• Complementarity of metrics: Ablation studies show single-metric optimization yields inferior results compared to the combined approach.

Notably, MorphAgent adapts efficiently across domain shifts, with negligible loss in task performance. Scalability analyses further indicate stable accuracy as agent count increases, though interaction rounds grow sublinearly. Computational overhead from continuous profile evaluation is highlighted as the main limitation.

Practical and Theoretical Implications

Horizontally scalable MAS frameworks are a prerequisite for deploying agents into unpredictable, open-world AI tasks. MorphAgent’s self-evolving, decentralized paradigm mitigates bottlenecks inherent in centralized orchestration and SOP-driven task planning. The use of dynamic role evolution fosters emergent swarm intelligence, robust specialization, and resilience against node failures. The framework is readily extensible to heterogeneous mixed-initiative teams and scenarios with shifting task requirements.

On the theoretical front, MorphAgent reifies emergent organization and adaptation within LLM-based MAS by operationalizing metrics for profile clarity, diversity, and task alignment. This direct mapping from swarming principles to algorithmic instantiations highlights the feasibility of robust MAS design absent centralized control.

Speculation on Future Developments

MorphAgent sets the foundation for broader work in peer-to-peer optimization, agent-oriented planning, and autonomous workflow design for AI teams. Future work may focus on reducing computational cost of metric evaluation, integrating additional behavioral metrics, expanding profile representation modalities (beyond text), and developing analytical frameworks for monitoring emergent collaboration structures at scale.

Conclusion

MorphAgent introduces a decentralized, self-adaptive framework for multi-agent collaboration, supporting dynamic role evolution and robust teamwork via profile-driven optimization. Empirical results demonstrate strong resilience and superior performance under failure and domain shift conditions. This paradigm offers a concrete pathway for the development of versatile, scalable, and autonomous MAS in complex real-world settings.