Self-Evolving AI Agents: A Comprehensive Survey of the New Paradigm

The Rise of Self-Evolving AI Agents

The field of artificial intelligence is undergoing a fundamental paradigm shift. While large language models (LLMs) have demonstrated remarkable capabilities in planning, reasoning, and natural language understanding, most existing agent systems share a critical limitation: they rely on manually crafted configurations that remain static after deployment. In a world where user intents shift, task requirements change, and external tools evolve, static agents are increasingly inadequate.

This comprehensive survey, authored by researchers from eight leading universities including Glasgow, Sheffield, Cambridge, the National University of Singapore, and UCL, introduces the concept of self-evolving AI agents — a new class of agent systems capable of autonomous adaptation and continuous self-improvement. The paper bridges the static capabilities of foundation models with the continuous adaptability required by lifelong agentic systems, offering the most systematic treatment of this emerging paradigm to date.

The timing of this survey is significant. As organizations worldwide deploy AI agents in increasingly complex real-world scenarios — from enterprise applications documented by McKinsey to cutting-edge research systems — the limitations of static agent architectures become more apparent. The ability for agents to autonomously evolve their capabilities based on experience represents the next frontier in AI development, with profound implications for how we design, deploy, and govern autonomous systems.

The Three Laws of Self-Evolving AI Agents

In a conceptual framework inspired by Isaac Asimov’s Three Laws of Robotics, the survey proposes Three Laws of Self-Evolving AI Agents that establish a hierarchical safety framework for autonomous agent evolution:

First Law — Endure (Safety Adaptation): Self-evolving AI agents must maintain safety and stability during any modification. This is the supreme constraint — no evolutionary improvement can compromise the agent’s fundamental safety properties.

Second Law — Excel (Performance Preservation): Subject to the First Law, self-evolving AI agents must preserve or enhance existing task performance. Evolution must not create regressions in capability. The agent should only become better, never worse.

Third Law — Evolve (Autonomous Evolution): Subject to the First and Second Laws, self-evolving AI agents must be able to autonomously optimize their internal components in response to changing tasks, environments, or resources. Autonomous adaptation is the goal, but always within safety and performance constraints.

This hierarchical structure ensures that safety always takes precedence over performance, and performance always takes precedence over unbounded evolution. It provides a principled foundation for designing self-evolving systems that are both capable and trustworthy — a crucial consideration as AI agents take on increasingly autonomous roles in society and enterprise.

From MOP to MASE: The Four Paradigms of AI Evolution

The survey characterizes the emergence of self-evolving AI agents as part of a broader four-stage paradigm shift in LLM-based system development. Each paradigm builds on the previous one, representing increasing levels of autonomy, adaptability, and sophistication.

Model Offline Pretraining (MOP) is the foundational stage — training models on large-scale static corpora and deploying them in a frozen state. Model Online Adaptation (MOA) introduces post-deployment updates through techniques like supervised fine-tuning, LoRA adapters, and RLHF, allowing models to improve from labels, ratings, or instruction prompts. These first two paradigms focus on single-model improvement.

Multi-Agent Orchestration (MAO) extends beyond single models to coordinate multiple LLM agents that communicate and collaborate via message exchange, debate prompts, or function calling to solve complex tasks without modifying underlying model parameters. The frontier paradigm — Multi-Agent Self-Evolving (MASE) — introduces a lifelong, self-evolving loop where a population of agents continually refines their prompts, memory, tool-use strategies, and interaction patterns based on environmental feedback and meta-rewards.

This evolution from MOP to MASE represents a fundamental shift from static, manually configured architectures to adaptive, data-driven systems that evolve in response to changing requirements. The implications are profound for the AI industry, as documented in research from leaders like DeepSeek’s reinforcement learning advances and Google’s Gemini 2.5 architecture.

The Unified Conceptual Framework

The survey introduces a unified conceptual framework that abstracts the feedback loop underlying self-evolving agentic systems. This framework identifies four key components: System Inputs (tasks, goals, and contextual information), the Agent System (the collection of agents and their internal components), the Environment (the external world the agents interact with), and Optimisers (the mechanisms that drive evolution based on feedback).

The framework serves as a foundation for understanding and comparing different self-evolution strategies. The feedback loop works as follows: the agent system receives inputs and interacts with the environment, which generates signals (rewards, corrections, performance metrics). The optimisers process these signals and modify the agent system’s components to improve future performance. This continuous cycle is what distinguishes self-evolving agents from static systems.

What makes this framework particularly valuable is its generality. Whether we are discussing a single agent optimizing its prompts based on user feedback, or a population of agents co-evolving their communication protocols in a multi-agent environment, the same abstract loop applies. This unified view enables researchers to identify commonalities, transfer techniques across domains, and identify gaps in the current literature.

Foundation Model Evolution

The survey reviews techniques for enhancing the underlying LLM to improve core capabilities such as planning, reasoning, and tool use. Foundation model evolution represents the most fundamental form of agent improvement — changes at this level affect every capability built on top of the model.

Key techniques include self-play and iterative self-improvement, where agents generate their own training data through interaction and use it to fine-tune their foundation models. Reinforcement learning approaches, particularly RLHF (Reinforcement Learning from Human Feedback) and its variants like DPO (Direct Preference Optimization), allow models to align their behavior with desired outcomes based on feedback signals.

The survey also covers parameter-efficient adaptation methods like LoRA (Low-Rank Adaptation) and prefix tuning, which enable model updates without the computational cost of full fine-tuning. These techniques are particularly relevant for self-evolving agents, as they allow rapid, targeted adaptations based on domain-specific feedback while preserving the model’s broad capabilities.

Prompt and Memory Optimization

Beyond modifying the foundation model itself, self-evolving agents can optimize their prompts and memory systems to improve performance. Prompt optimization involves automatically refining the instructions, examples, and context provided to the LLM to elicit better responses for specific tasks.

The survey catalogs various approaches: gradient-free prompt optimization that searches the space of natural language prompts, automatic prompt engineering that generates and evaluates candidate prompts, and meta-prompt strategies that use the LLM itself to improve its own prompts. These techniques allow agents to adapt their behavior without any model parameter changes — a critical capability when model fine-tuning is impractical or undesirable.

Memory optimization addresses how agents store, organize, and retrieve information from past interactions. Effective memory systems allow agents to learn from experience, avoid repeating mistakes, and build on previous successes. The survey covers episodic memory (specific past events), semantic memory (general knowledge), and procedural memory (learned workflows), along with techniques for each type. As AI systems scale, memory management becomes a critical bottleneck — an agent that can intelligently evolve its memory strategies gains a significant advantage over static alternatives.

Tool and Workflow Evolution

One of the most practically significant areas of agent evolution is tool optimization. Self-evolving agents can learn to create new tools, improve existing ones, and optimize their selection and usage strategies based on experience. This is particularly relevant as the ecosystem of available tools (APIs, databases, web services) continuously changes.

The survey reviews techniques for automatic tool creation, where agents generate new tools (typically as code functions) to address capability gaps they discover during task execution. Tool selection optimization improves how agents choose among available tools for specific tasks, while tool composition strategies help agents learn to chain multiple tools together in novel ways to solve complex problems.

Workflow optimization addresses the higher-level structure of how agents decompose and execute tasks. This includes evolving task decomposition strategies, planning algorithms, and error recovery procedures. Self-evolving agents can learn from failed task executions to develop more robust workflows, adapting their approach based on the specific characteristics of different task types and environments. The computational infrastructure documented in NVIDIA’s latest report provides the hardware foundation that makes these sophisticated evolution processes feasible at scale.

Multi-Agent Communication and Collaboration

In multi-agent systems, evolution extends beyond individual agent capabilities to encompass inter-agent communication protocols and collaboration strategies. The survey explores how populations of agents can co-evolve their interaction patterns to achieve better collective performance — a particularly exciting frontier in AI research.

Key areas include the evolution of debate and discussion protocols, where agents learn more effective ways to exchange information, challenge each other’s reasoning, and reach consensus. Role specialization evolution allows agents within a system to develop distinct expertise areas that complement each other, improving the system’s overall capability to handle diverse tasks.

The survey also covers communication efficiency optimization, where agents learn to convey maximum information with minimum overhead — a critical concern as multi-agent systems scale to larger numbers of participants. Techniques like selective information sharing, summarization strategies, and adaptive communication frequencies allow self-evolving multi-agent systems to maintain coordination without being overwhelmed by communication costs.

Domain-Specific Evolution Strategies

The survey dedicates significant attention to domain-specific evolution strategies developed for specialized fields where agent behavior and optimization objectives are tightly coupled with domain constraints. Three domains receive particular focus: biomedicine, programming, and finance.

In biomedicine, self-evolving agents must navigate strict safety requirements, evidence-based reasoning standards, and rapidly evolving medical knowledge. Evolution strategies in this domain emphasize safe exploration, rigorous validation of proposed actions, and integration with authoritative medical databases. The stakes of getting things wrong — patient safety — make the First Law (Endure) particularly critical.

In programming, agents can leverage the unique advantage of automated testing: code either works or it doesn’t. This creates a natural feedback loop that enables rapid evolution of coding strategies, debugging approaches, and code generation techniques. Self-evolving programming agents can learn from execution errors, test failures, and code review feedback to continuously improve their development capabilities.

In finance, agents must evolve within strict regulatory constraints while adapting to rapidly changing market conditions. Evolution strategies focus on risk management, compliance awareness, and the ability to process and act on real-time market data. The combination of high-frequency decision-making and regulatory oversight creates unique challenges for self-evolution that require careful balancing of adaptability and constraint compliance. The regulatory and compliance frameworks seen in major corporations illustrate the kinds of constraints financial AI agents must navigate.

Safety, Ethics, and Evaluation Challenges

The survey provides a dedicated discussion on safety, ethics, and evaluation considerations for self-evolving agentic systems — areas the authors identify as critical to ensuring effectiveness and reliability. As agents gain the ability to modify their own behavior autonomously, the potential for unintended consequences grows significantly.

Safety challenges include preventing agents from evolving behaviors that circumvent intended constraints, ensuring that evolutionary processes don’t amplify biases present in training data or environments, and maintaining human oversight as agents become more autonomous. The Three Laws framework provides a theoretical foundation, but practical implementation remains an active research area.

Evaluation challenges are equally significant. How do you assess whether an agent’s self-modifications actually improve its capabilities in a meaningful and generalizable way? The survey reviews evaluation benchmarks, metrics, and methodologies, noting that many existing evaluation frameworks are designed for static systems and may not adequately capture the dynamics of self-evolving agents. New evaluation paradigms that can assess agents over time, across changing environments, and under distribution shift are needed.

The survey concludes by identifying key open research directions: developing more robust safety guarantees for self-evolving systems, creating evaluation frameworks that capture long-term evolution dynamics, enabling efficient evolution in resource-constrained environments, and establishing governance frameworks that balance autonomy with accountability. The paper is available in full on arXiv, with an accompanying resource collection at GitHub. For broader context on AI safety and governance, see NIST’s AI framework.