PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems

Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, Zhaochun Ren, Nikos Aletras, Xi Wang, Han Zhou, Zaiqiao Meng

arXiv:2508.07407v2cs.AIcs.CLcs.MA

TL;DR

Static, manually configured agent systems struggle to adapt as requirements and environments change. This survey introduces a feedback-loop framework and synthesizes self-evolution techniques, evaluation approaches, and safety constraints for more adaptive agentic systems.

  • Problem

    Static, manually configured LLM-based systems have limited ability to respond to changing requirements and environments.

  • Method

    The survey introduces a four-component feedback-loop framework and systematically reviews techniques for evolving agent systems through environmental feedback and optimisation.

  • Results

    The survey synthesizes optimisation techniques across agent components, domain-specific strategies, and evaluation methodologies for adaptive and resilient agentic systems.

  • Takeaways & Limitations

    Self-evolving agents mark a shift from static, human-configured models toward dynamic ecosystems guided by safety, performance, and adaptability constraints.

  • Takeaways & Limitations

    Current evaluations are mostly snapshot-based, leaving longitudinal benchmarks for safety, alignment, and robustness across an agent's lifecycle as an open challenge.

Abstract

from arXiv · show

Recent advances in large language models have sparked growing interest in AI agents capable of solving complex, real-world tasks. However, most existing agent systems rely on manually crafted configurations that remain static after deployment, limiting their ability to adapt to dynamic and evolving environments. To this end, recent research has explored agent evolution techniques that aim to automatically enhance agent systems based on interaction data and environmental feedback. This emerging direction lays the foundation for self-evolving AI agents, which bridge the static capabilities of foundation models with the continuous adaptability required by lifelong agentic systems. In this survey, we provide a comprehensive review of existing techniques for self-evolving agentic systems. Specifically, we first introduce a unified conceptual framework that abstracts the feedback loop underlying the design of self-evolving agentic systems. The framework highlights four key components: System Inputs, Agent System, Environment, and Optimisers, serving as a foundation for understanding and comparing different strategies. Based on this framework, we systematically review a wide range of self-evolving techniques that target different components of the agent system. We also investigate domain-specific evolution strategies developed for specialised fields such as biomedicine, programming, and finance, where optimisation objectives are tightly coupled with domain constraints. In addition, we provide a dedicated discussion on the evaluation, safety, and ethical considerations for self-evolving agentic systems, which are critical to ensuring their effectiveness and reliability. This survey aims to provide researchers and practitioners with a systematic understanding of self-evolving AI agents, laying the foundation for the development of more adaptive, autonomous, and lifelong agentic systems.

1 Introduction

Self-evolving AI agents address the limits of manually configured, static systems by adapting through interaction data and environmental feedback. This survey frames the field with guiding laws, a unified feedback-loop framework, systematic technique reviews, and discussions of domain-specific strategies, evaluation, safety, ethics, and open challenges.

  • Most deployed agents retain manually designed configurations and static architectures despite continuously changing tasks, users, tools, and information sources.These conditions motivate systems capable of autonomous adaptation and continuous self-improvement.
  • The survey formalises three laws requiring safety and stability, preserved or enhanced performance, and autonomous optimisation under changing tasks, environments, or resources.These laws are presented as guiding principles for safe and effective self-evolution.
  • The field progresses from Model Offline Pretraining and Model Online Adaptation through Multi-Agent Orchestration toward Multi-Agent Self-Evolving systems.This trajectory moves from static, frozen foundation models toward autonomous systems adapting to changing requirements and environments.
  • Its unified framework abstracts the evolution loop into System Inputs, Agent System, Environment, and Optimisers for understanding and comparing approaches.The survey reviews techniques targeting foundation models, prompts, memory, tools, workflows, and inter-agent communication.
  • The review spans single-agent, multi-agent, and domain-specific optimisation, alongside evaluation, safety, ethical considerations, and future research directions.The survey identifies these considerations as important for effective, safe, responsible, adaptive, and sustainable agentic systems.

2 Foundation of AI Agent Systems

AI agents combine a foundation model with supporting modules for perception, planning, memory, and tool use, while multi-agent systems coordinate specialised agents through structured workflows. The survey motivates self-evolving systems that reduce reliance on handcrafted configurations by adapting prompts, memory, tools, and interaction topologies in response to changing environments.

  • AI Agents: AI agents autonomously perceive inputs, reason about goals, and interact with environments to complete tasks.
  • AI Agent Components: A foundation model serves as the core reasoning engine, while perception, planning, memory, and tool modules support autonomous decision-making and execution.Perception interprets environmental information; planning decomposes tasks; memory retains experience; tools extend interaction with real-world environments.
  • AI Agent Components: Memory supports context-aware reasoning through short-term task context and long-term storage of accumulated knowledge, experiences, or reusable information.Retrieval-augmented generation can retrieve relevant long-term memory and incorporate it into the LLM’s input context.
  • Multi-Agent Systems: Multi-agent systems coordinate autonomous agents in a shared environment to achieve goals beyond a single agent’s capabilities.Agent topology determines information flow, collaboration strategies, task distribution, and execution across the workflow.
  • System Architecture: Multi-agent architectures range from static hierarchies to centralised manager-follower designs, with central coordination balancing global planning and task execution.Central nodes can create performance bottlenecks and single-point-of-failure vulnerabilities.
  • Lifelong Self-Evolving Systems: Self-evolving systems close the deployment–improvement loop by autonomously refining prompts, memory, tool-use strategies, and inter-agent topologies from environmental feedback and meta-rewards.This shifts agents from static executors toward configurations that generate, evaluate, and refine themselves as tasks, domains, resources, and constraints change.

3 A Conceptual Framework of MASE

The survey frames self-evolving agent systems as iterative optimisation loops connecting system inputs, an agent system, an environment, and optimisers. These components define the task, execute it, generate feedback, and update agent configurations until performance or convergence criteria are met.

  • Framework Overview: The framework abstracts self-evolving agents as a feedback loop with four components: system inputs, agent system, environment, and optimisers.It is intended to provide a generalisable basis for understanding and comparing agent evolution approaches.
  • Iterative Evolution: Each cycle redeploys the updated agent system, optionally refines system inputs, and stops when a performance threshold or convergence criterion is reached.The loop can update LLM parameters, prompts, system structure, or training examples.
  • System Inputs: System inputs specify task requirements, constraints, and available data, with task-level and instance-level optimisation as two settings.Task-level inputs may include T and D_train, whereas instance-level inputs may include (x, y, C).
  • Agent System: The agent system is the optimised decision-making core, comprising a single agent or collaborating agents and components such as LLMs, prompts, memory, and tools.Methods may optimise one component or jointly optimise multiple components.
  • Environment: The environment supplies the operating context, receives agent actions, and produces task-specific feedback metrics such as accuracy, F1, or success rate.It can range from benchmark datasets to dynamic real-world settings.
  • Optimisers: Optimisers search configuration space S for an agent A* that maximises evaluation function O(A; I) under the supplied inputs.Search algorithms may include heuristics, gradient descent, Bayesian optimisation, or Monte Carlo Tree Search.

4 Single-Agent Optimisation

Single-agent optimisation improves one agent by selecting appropriate search spaces and optimisation strategies. The survey organises approaches around the agent components they target, including LLM behaviour, prompts, memory, and tools.

  • Overview: Single-agent optimisation focuses on updating one agent system through a chosen search space and optimisation algorithm.The optimiser must determine which capability and component to improve.
  • Overview: The survey categorises single-agent approaches by LLM behaviour, prompt, memory, and tool optimisation.These categories determine both the search space and optimisation methods.
  • Prompt Optimisation: Prompt optimisation adapts instructions to guide the LLM toward more accurate and task-relevant outputs.It is one of the four major single-agent optimisation categories.
  • Memory Optimisation: Memory optimisation improves an agent’s ability to store, retrieve, and reason over historical information or external knowledge.The category targets how agents use retained information during task execution.
  • Tool Optimisation: Tool optimisation improves an agent’s use of existing tools or enables it to create or configure tools for complex tasks.The approach targets tool-use effectiveness rather than only the underlying model.

4.1 LLM Behaviour Optimisation

LLM behaviour optimisation addresses the gap between language fluency and complex reasoning in agent systems. Reviewed approaches either update model parameters through training or increase reasoning effort at test time using feedback and search.

  • Overview: LLM behaviour optimisation targets planning and reasoning because these capabilities underpin single-agent task execution.Methods broadly use training-based parameter updates or test-time computation.
  • Training-Based Optimisation: Supervised fine-tuning trains agents on annotated reasoning steps that map questions through intermediate processes to final answers.Reasoning trajectories may come from agent rollouts or stronger teacher models.
  • Training-Based Optimisation: Reinforcement learning treats reasoning as sequential decision-making and rewards correct or high-quality reasoning paths.Preference-based optimisation can use outcome, test-case, or process-reward-model signals.
  • Test-Time Optimisation: Test-time compute enables inference-time reasoning refinement when training resources are constrained or API models cannot be fine-tuned.Increasing the inference budget allows models to “think longer.”
  • Test-Time Optimisation: Feedback-based strategies adjust generated behaviour using verifier scores, which may evaluate final outcomes or reasoning processes.The feedback guides refinement without necessarily changing model parameters.
  • Test-Time Optimisation: Search-based strategies generate multiple reasoning trajectories and use critics or verifiers to select promising solutions.CoT-SC is described as a best-of-N approach that samples multiple reasoning paths.

4.2 Prompt Optimisation

Prompt optimisation responds to LLM sensitivity to wording, formatting, and ordering by automatically searching for higher-quality prompts. The survey covers edit-based, generative, text-gradient, and evolutionary strategies.

  • 4.2 Prompt Optimisation: Minor prompt changes can significantly alter LLM behaviour, motivating automatic search for robust and generalisable prompts.Prompts encode instructions, demonstrations, and contextual information.
  • 4.2 Prompt Optimisation: Prompt optimisation methods are grouped into edit-based, generative, text gradient-based, and evolutionary approaches.The categories differ in how they navigate prompt space and identify candidates.
  • 4.2.1 Edit-Based Prompt Optimisation: Edit-based methods perform local token insertion, deletion, or substitution on human-written prompts.They treat optimisation as local search while preserving core prompt semantics.
  • 4.2.2 Generative Prompt Optimisation: Generative methods use LLMs to create entirely new prompts conditioned on base prompts and optimisation signals.Compared with local edits, they explore broader prompt-space regions and produce more diverse candidates.
  • 4.2.2 Generative Prompt Optimisation: Generative prompt optimisation can use rewriting rules, examples, dataset descriptions, prior scores, or meta-prompts as guidance.These signals steer the generation of improved prompt candidates.
  • 4.2.2 Generative Prompt Optimisation: Some generative approaches train a policy model with reinforcement learning to iteratively refine prompts from failed-case analysis.Retroformer is cited as an example that summarises failure root causes.
  • 4.2.3 Text Gradient-Based Prompt Optimisation: Text-gradient methods generate natural-language critiques instead of numerical parameter gradients, then use them to update prompts.The approach depends on how feedback is generated and applied.
  • 4.2.4 Evolutionary Prompt Optimisation: Evolutionary methods maintain populations of candidate prompts and refine them through mutation, crossover, and selection.They treat prompt optimisation as an evolutionary process.

4.3 Memory Optimisation

Memory optimisation addresses context limits and forgetting by organising short-term coherence and long-term persistence, with dynamic control over storage, updating, and retrieval.

  • Memory Optimisation Objectives: Memory optimisation targets short-term coherence and long-term retrieval across sessions.This perspective emphasises dynamic memory control rather than static internal or external formats.
  • Short-term Memory Optimisation: Short-term memory methods compress, summarise, or selectively retain recent dialogue, reasoning traces, and task-relevant context.Examples include recursive summarisation, dialogue-level memory, sparse attention, and dynamic context filtering.
  • Short-term Memory Optimisation: Short-term memory improves local coherence and context efficiency but cannot alone preserve knowledge across sessions or support long-horizon generalisation.Long-term memory mechanisms are therefore complementary rather than optional for persistent use.
  • Long-term Memory Optimisation: Long-term memory provides persistent storage for factual knowledge, task histories, user preferences, and interaction trajectories across sessions.Retrieval-augmented generation incorporates relevant external memory into the reasoning process.
  • Long-term Memory Optimisation: Graph-based and cognitively inspired systems improve memory interpretability, conversational dependency capture, and retrieval guidance.Examples include hippocampus-inspired indexing, graph structures, and symbolic memory systems.
  • Memory Control: Inference-time memory control determines what, when, and how agents store, update, or discard information.Representative systems dynamically select heterogeneous memory segments and support continuous online or offline updates.

4.4 Tool Optimisation

Tool optimisation improves how agents understand, select, and execute external tools, while also adapting tool documentation, reasoning procedures, and the toolset itself.

  • Overview: Tool optimisation follows two directions: improving agent interaction with tools and optimising the tools themselves.The first includes training, prompting, and reasoning; the second adapts or generates tools for task-specific execution.
  • Training-Based Tool Optimisation: Training-based methods update LLM parameters because pretrained models lack inherent knowledge of tool invocation and output interpretation.Supervised fine-tuning uses high-quality trajectories containing queries, reasoning steps, tool calls, and final answers.
  • Training-Based Tool Optimisation: High-quality tool-use trajectories supervise planning, execution, and incorporation of tool results into reasoning.MCTS and multimodal synthesis pipelines are used to collect or generate stronger trajectories.
  • Training-Based Tool Optimisation: Multi-turn tool calling remains challenging when function calls are complex, dependencies are long-term, or information is missing.Magnet synthesises query and executable-call sequences and uses graphs to construct reliable multi-turn trajectories.
  • Reinforcement Learning for Tool Optimisation: Supervised fine-tuning is constrained by training-data quality and coverage, which can limit gains and generalisation to unseen tools or configurations.These limitations motivate reinforcement-learning approaches for tool optimisation.
  • Inference-Time Tool Optimisation: Inference-time methods optimise tool-related context or guide decision-making without modifying LLM parameters.They include prompt-based documentation refinement and structured reasoning at test time.
  • Prompt-Based Tool Optimisation: Jointly refining agent instructions and tool descriptions can reduce computational overhead and improve tool-use efficiency.The combined representation is treated as an optimised context.

5 Multi-Agent Optimisation

Multi-agent optimisation progresses from manually designed collaboration toward automated search over prompts, topologies, and backbone models, with workflows treated as evolving design spaces.

  • Overview: The field progresses from handcrafted architectures to prompt refinement, topology discovery, joint optimisation, and backbone training.These dimensions broaden what can be optimised in multi-agent collaboration.
  • Overview: Multi-agent workflow optimisation spans prompt, topology, unified, and LLM-backbone dimensions across core spaces, methods, and targets.Figure 6 organises these elements as a structured overview of multi-agent systems optimisation.
  • Manually Designed Workflows: Parallel workflows generate solutions concurrently and aggregate them, while hierarchical workflows decompose dependent subtasks across layers.Parallel generation with small LLMs can match or outperform single large LLMs, whereas hierarchical designs suit complex goal-driven tasks.
  • Manually Designed Workflows: Debate workflows use adversarial negotiation and arbitration to discuss and correct reasoning errors.Confidence-gated debate can trigger discussion selectively to reduce inference costs without hindering performance.
  • Self-Evolving Workflows: Self-evolving workflows automatically design, evaluate, and refine prompts, topologies, and collaboration strategies using performance feedback.They replace hard-coded configurations with search over possible workflow designs.
  • Self-Evolving Workflows: Workflow search uses reinforcement learning, Monte Carlo Tree Search, generative models, and evolutionary operators while balancing accuracy, efficiency, and safety.The optimisation objective has expanded beyond task performance alone.
  • Topology Optimisation: Topology optimisation treats communication structure as an optimisation target, separating executable code-level workflows from communication-graph topologies.This reframes topology as a medium that matters alongside prompt content.
  • Topology Optimisation: Code-level methods search executable programs or typed code graphs, while communication-graph methods optimise agent connections through learned edge structures.Examples use reinforcement learning, Monte Carlo Tree Search, or continuous edge probabilities to explore workflow spaces.

6 Domain-Specific Optimisation

Domain-specific optimisation adapts agent architectures and capabilities to specialised knowledge, data modalities, task structures, and operational constraints, with biomedical diagnosis and molecular discovery as examples.

  • Domain-Specific Challenges: Domain-specific agent systems require tailored optimisation because specialised domains impose distinct structures, knowledge bases, modalities, and operational constraints.The survey covers biomedicine, programming, scientific research, games, computer use, finance, and legal research.
  • Biomedical Optimisation: Biomedical optimisation aligns agent behaviour with procedural and operational requirements in medical diagnosis and molecular discovery.These are identified as two key biomedical application areas.
  • Medical Diagnosis: Medical diagnosis agents conduct diagnostic dialogues, ask clarifying questions, and generate hypotheses from incomplete or ambiguous patient information.Robust reasoning may require external knowledge bases and specialised medical tools.
  • Medical Diagnosis: Multi-agent diagnosis systems use simulation-driven or collaborative designs to model complex, multi-step clinical reasoning.Simulation-driven systems assign roles and learn diagnostic strategies through interaction in simulated medical environments.
  • Medical Diagnosis: Tool integration and multimodal reasoning improve diagnostic generalisability and clinical reliability through specialised tools, planning, and multimodal evidence.MMedAgent and MedAgent-Pro exemplify these directions.
  • Molecular Discovery: Molecular discovery agents address chemical structures, reaction pathways, and pharmacological constraints through tools, memory, and multi-agent collaboration.These systems target precise symbolic reasoning in chemically constrained tasks.
  • Molecular Discovery: Memory-enabled molecular systems record prior solutions, decompose chemical tasks, and retrieve structured subtasks for refinement.Modular reasoner, executor, evaluator, and memory roles support iterative workflows and multi-objective molecule evaluation.

6.2 Domain-Specific Optimisation in Programming

Programming-focused agent optimisation adapts agent behaviour, roles, workflows, and tool use to software-engineering requirements. The surveyed work centres on code refinement and debugging, using iterative feedback, collaboration, memory, and execution-aware reasoning.

  • Programming Optimisation: Programming-domain optimisation targets code refinement and debugging by aligning agent behaviour with established software-engineering workflows.The section identifies these as the two key application areas for domain-specific agent design.
  • Code Refinement: Code refinement improves code quality, structure, and correctness while preserving original functionality.Agent systems support this through self-improvement, collaborative workflows, and programming-tool integration.
  • Code Refinement: Experience-driven learning lets agents record and reuse solutions to previously encountered tasks through memory-enabled reasoning.AgentCoder and CodeAgent exemplify collaborative development workflows built around this approach.
  • Code Refinement: Adaptive code-refinement systems emphasise iterative feedback, modular design, and interactive reasoning.These principles are presented as essential for building adaptive agent-based systems.
  • Code Debugging: Code debugging requires fault localisation, execution-aware reasoning, and iterative correction that general-purpose LLMs typically lack.Domain-specific optimisation therefore aligns agent roles and workflows with structured human debugging practices and tool use.
  • Code Debugging: Modular debugging architectures separate responsibilities across agents, such as code generation and semantic validation.PyCapsule separates programmer and executor agents, while other systems use collaborative multi-stage pipelines.

6.3 Domain-Specific Optimisation in Financial and Legal Research

Financial and legal agent optimisation tailors architectures, reasoning strategies, collaboration, and tools to domain-specific procedures and constraints. Financial systems address volatile heterogeneous information, while legal systems emphasise rules, evidence, procedural realism, and interpretability.

  • Overview: Financial and legal optimisation adapts multi-agent architectures, reasoning strategies, and tool integration to domain-specific workflows.Reliable performance depends on modular design, collaborative interaction, and rule-grounded reasoning.
  • Financial Decision-Making: Financial decision-making requires reasoning over volatile markets while integrating numerical indicators, news sentiment, and expert knowledge.These requirements motivate architectures tailored to financial procedures and cognitive demands.
  • Financial Decision-Making: Financial reporting systems combine specialised sentiment modules with rule-based validators to improve robustness and compliance.Template-based frameworks additionally decompose report generation into agent-driven stages.
  • Legal Reasoning: Legal reasoning requires structured interpretation of rules, case-specific evidence analysis, and outputs consistent with institutional and judicial standards.Research therefore uses multi-agent systems tailored to legal procedural and interpretive requirements.
  • Legal Reasoning: AgentCourt uses reflective self-play to refine lawyer-agent strategies, improving debate quality and procedural realism.The approach directly targets the interaction and realism of legal debate.
  • Legal Reasoning: LegalGPT and AgentsCourt support interpretable, rule-aligned reasoning through legal chains of thought, courtroom simulation, and knowledge augmentation.Their outputs are grounded in codified rules and case precedents.

7 Evaluation

Agent evaluation is expanding from task completion toward reasoning, generalisation, coordination, tool use, safety, and lifecycle behaviour across diverse environments. LLM- and agent-based judges offer scalable assessment, but dynamic settings, bias, generalisation, and longitudinal safety remain unresolved concerns.

  • Evaluation Scope: Evaluation increasingly measures task completion alongside reasoning quality, generalisation, and safety or alignment compliance.The survey frames evaluation as dynamic feedback rather than a static endpoint.
  • Specialised Benchmarks: Tool-use benchmarks assess the correctness and efficiency of external API and function calls in simulated or sandboxed environments.ToolBench, API-Bank, MetaTool, and ToolQA exemplify this evaluation focus.
  • Specialised Benchmarks: Web benchmarks test navigation, interface adaptability, information extraction, and textual-visual integration across simulated and live environments.Intermediate sub-goal metrics and robustness assessments are increasingly included.
  • Specialised Benchmarks: Multi-agent and generalist benchmarks evaluate collaboration, competition, decentralised coordination, and competence across heterogeneous tasks.These evaluations expose difficulties in aggregating metrics and risks of overfitting to narrow scenarios.
  • Specialised Benchmarks: GUI, multimodal, and domain-focused benchmarks require complex action sequences combining language understanding, visual perception, tool use, and domain norms.Examples span operating systems, coding, data science, enterprise productivity, and scientific research.
  • LLM-as-a-Judge: LLM-as-a-Judge uses structured prompts for pointwise scoring or pairwise comparison, offering scalable assessment but remaining sensitive to prompt-induced bias.Single-step output evaluation may also miss reasoning depth in multi-step processes.
  • Agent-as-a-Judge: Agent-as-a-Judge evaluates complete reasoning trajectories using multi-step reasoning, state management, tools, intermediate artefacts, and reasoning graphs.On DevAI, these evaluations aligned more closely with expert judgments and reduced review time and cost relative to manual assessment.
  • Agent-as-a-Judge: Agent-as-a-Judge adds implementation complexity and faces generalisation challenges beyond code-generation domains.Current work aims to improve adaptability and simplify deployment across broader tasks.

8 Challenges and Future Directions

The survey identifies safety, instability, evaluation, efficiency, transferability, multimodality, and tool evolution as central obstacles to lifelong self-evolving agents. It proposes evolution-aware benchmarks, autonomous tool and environment interaction, domain-aware methods, and optimisation that jointly considers performance and resources.

  • Challenges: Current agent evolution pipelines often prioritise task metrics over safety, risking unintended behaviour, privacy breaches, and misaligned objectives.The survey calls for evolution-aware audits, adaptive licences, safety sandboxes, and legal protocols.
  • Challenges: Reward optimisation can become unstable because intermediate-step reward models face scarce data, noisy supervision, and inconsistent feedback.Small perturbations in inputs or update rules may undermine trustworthiness.
  • Challenges: Scientific and specialised domains lack reliable, uncontested ground truth, complicating the construction of trustworthy optimisation feedback.This is highlighted for domains such as biomedicine and law.
  • Challenges: Large-scale multi-agent optimisation can improve task performance while increasing computational cost, latency, and instability.The unresolved challenge is to trade off effectiveness against efficiency explicitly.
  • Challenges: Optimised prompts and agent topologies can be brittle across LLM backbones with different reasoning abilities, limiting scalability and reuse.This transferability problem affects production deployment.
  • Challenges: Most optimisation algorithms remain text-only, although real-world agents must process multimodal inputs and reason in spatially grounded or continuous environments.The survey identifies internal world models and perceptual-temporal reasoning as requirements.
  • Future Directions: Current methods generally assume fixed toolsets, leaving autonomous tool discovery, adaptation, and co-evolution unresolved.This motivates research on agents that can adaptively select, compose, or create tools.
  • Future Directions: Open-ended interactive simulation platforms could let agents iteratively receive feedback and refine prompts, memory, tools, and workflows through closed-loop optimisation.These environments are proposed as infrastructure for fully autonomous self-evolution.

9 Conclusions

The survey frames self-evolving AI agents as a progression toward dynamic, autonomous ecosystems and formalises their evolution through a unified feedback-loop framework. It also proposes safety and performance guardrails while identifying research advances needed for real-world deployment.

  • Evolution trajectory: The survey presents a four-stage trajectory from MOP and MOA through MAO to MASE, marking a shift from static models to autonomous ecosystems.MASE denotes Multi-Agent Self-Evolving systems.
  • Conceptual framework: Its framework models continual agent improvement through four components: Inputs, Agent System, Objectives, and Optimisers.These components abstract the feedback loop underlying agent evolution.
  • Survey scope: The survey reviews optimisation techniques across agent components, domain-specific strategies, and evaluation methodologies for adaptive and resilient agentic systems.The reviewed strategies span agent-system optimisation and specialised domains.
  • Safety principles: The Three Laws—Endure, Excel, and Evolve—are proposed as practical design constraints for safe, effective, and aligned lifelong self-improvement.They function as guardrails for the MASE paradigm and continuous open-ended self-improvement.
  • Future directions: Progress toward real-world self-evolving agents requires scalable optimisation, lifelong evaluation, safe heterogeneous-agent coordination, and adaptation to unforeseen domains.The paper identifies these as needed breakthroughs for dynamic environments.
  • Outlook: The survey calls for an ecosystem of self-evolving agents that can live, learn, and last while remaining autonomous, resilient, and trustworthy.It positions principled self-evolution as a basis for lifelong agentic systems.
Loading 2508.07407v2…