Beyond Single-Agent: Survey of Collaboration, Attribution, and Self‑Evolution in LLM Multi‑Agents

In the past two years, large‑language‑model (LLM) agents have evolved from simple input‑output modules to systems that can understand tasks, decompose steps, use tools, retain memory, and adapt behavior based on feedback. When a single agent cannot handle a complex task, researchers organize multiple agents to cooperate, which introduces new scalability challenges.

The survey proposes the LIFE progression, a complete observation framework for LLM‑based multi‑agent systems. It consists of four consecutive phases: Individual Intelligence, Multi‑Agent Collaboration, Failure Attribution, and Self‑Evolution.

1. Individual Intelligence

Modern LLM agents embed reasoning, memory, planning, and tool‑use mechanisms. Reasoning enables handling complex instructions; memory allows leveraging historical context; planning supports long‑term task decomposition; tool use extends the agent beyond the model’s internal knowledge.

2. Multi‑Agent Collaboration

Collaboration is the core of multi‑agent systems. Existing research focuses on four dimensions: role (defining each agent’s responsibility), communication (information flow), scheduling (task progression), and interaction patterns (task‑specific protocols). Different applications—code generation, scientific discovery, web navigation, complex QA, or game environments—require distinct collaboration designs, so a single fixed workflow cannot cover all scenarios.

Without stable individual capabilities, collaboration can amplify errors, increase communication overhead, and make system behavior unpredictable. Early mis‑judgments, incomplete information transfer, or faulty tool calls can propagate through the organization, altering the entire task trajectory.

3. Failure Attribution

In single‑agent settings, failures can often be traced to a clear input‑output mismatch. In multi‑agent systems, failures usually involve multiple interdependent steps. The survey emphasizes diagnosing *where* a failure occurs, *which* agents are involved, and *whether* the root cause lies in insufficient ability, role design, communication, scheduling, or environment interaction, as well as how the error spreads internally.

Many studies focus on constructing collaboration pipelines and boosting final performance, but few address post‑failure diagnostics. Without attribution, system improvements become blind trial‑and‑error, as poor performance does not directly indicate whether to modify the model, prompts, role allocation, communication protocol, or overall organization.

4. Self‑Evolution

Attribution asks *what* went wrong; self‑evolution asks *how* the system can become better. Existing reflective mechanisms—summarizing failure reasons and adjusting prompts—are insufficient for multi‑agent contexts because improvement targets include not only individual outputs but also the system’s structural components.

Agentic Self‑Evolution : updates internal components of a single agent (prompts, memory, parameters) to improve stability.

Systemic Self‑Evolution : modifies system‑level elements such as communication topology, agent composition, or shared memory, enabling collaborative patterns to adapt to tasks and feedback.

Meta Self‑Evolution : leverages accumulated design experience or generative models to automatically generate more suitable multi‑agent architectures for new tasks.

Thus, self‑evolution in multi‑agent systems resembles a systemic adjustment: based on task performance and failure feedback, the system continuously refines its behavior, structure, and collaboration mechanisms.

5. The LIFE Framework

The LIFE framework integrates the four phases into a single lifecycle: Individual Intelligence provides the foundation; Multi‑Agent Collaboration introduces system‑level complexity; Failure Attribution makes error processes analyzable; Self‑Evolution turns diagnostics into continual improvement.

6. Future Outlook

To achieve reliable long‑term deployment, the survey identifies several research directions: (1) richer evaluation metrics that cover communication efficiency, role contribution, error propagation, environmental adaptability, and stability; (2) dynamic collaboration structures that can reconfigure roles, communication paths, and scheduling on‑the‑fly; (3) tighter attribution‑to‑repair loops that guide concrete system adjustments; and (4) controllable self‑evolution mechanisms that balance performance gains with safety, cost, and alignment considerations.

The authors provide the paper link (https://arxiv.org/abs/2605.14892) and the project repository (https://github.com/mira-ai-lab/awesome-mas-life) for further exploration.