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

• Static, manually configured agent systems struggle to scale as user intents, task requirements, tools, and information sources change after deployment.Manual reconfiguration is time-consuming, labour-intensive, and difficult to scale in dynamic real-world settings.Source: §1 · p. 2
• Self-evolving AI agents autonomously adapt and continuously improve by optimising internal components through environmental interaction while preserving safety and performance.They bridge foundation models with lifelong learning agentic systems and respond to changing tasks, contexts, and resources.Source: §1 · p. 2
• The proposed Three Laws require agents to maintain safety and stability, preserve or enhance task performance, and autonomously optimise only subject to those constraints.These principles are formalised as Endure, Excel, and Evolve, respectively.Source: §1 · p. 3
• LLM-centric learning progresses from static Model Offline Pretraining and post-deployment Model Online Adaptation through Multi-Agent Orchestration to lifelong Multi-Agent Self-Evolving systems.MASE continually refines prompts, memory, tool-use strategies, and interaction patterns using environmental feedback and meta-rewards.Source: §1 · p. 3
• Existing surveys insufficiently cover agent self-evolution and continual adaptation, leaving a literature gap concerning lifelong, autonomous agentic systems.Prior surveys mainly address general architectures, individual components, evaluation, or domain-specific applications.Source: §1 · p. 4
• This survey introduces a four-component feedback-loop framework and systematically reviews evolution techniques, domain-specific strategies, evaluation, safety, ethics, challenges, and research directions.The framework comprises System Inputs, Agent System, Environment, and Optimisers, supporting comparison across single-agent, multi-agent, and domain-specific settings.Source: §1 · pp. 4–5

2.1 AI Agents

• AI agents autonomously perceive inputs, reason about goals, and interact with environments, typically using a foundation model alongside perception, planning, memory, and tool-use modules.Planning may be linear, iterative, or branching, while memory supports context-aware and long-term reasoning.Source: §2.1 · pp. 6–7

2.2 Multi-Agent Systems

• Multi-agent systems coordinate autonomous agents through topologies and workflows to achieve shared goals beyond individual capabilities, enabling specialization, parallel execution, robustness, scalability, and collaborative refinement.Their collective performance can exceed the sum of individual capabilities, with frameworks such as CAMEL and AutoGen reducing orchestration overhead.Source: §2.2 · p. 7

2.2.1 System Architecture

• Multi-agent architectures span hierarchical, centralized, and decentralized structures, trading modular control and global coordination against bottlenecks, single-point failures, synchronization challenges, and collaboration costs.Centralized designs risk coordinator bottlenecks, whereas decentralized designs improve robustness but complicate information synchronization and security.Source: §2.2.1 · p. 8

2.2.2 Communication Mechanisms

• Multi-agent communication balances structured formats’ machine readability and efficiency with natural language’s contextual richness, while protocols such as A2A, ANP, MCP, and Agora improve interoperability and negotiation.Natural language remains expressive but can introduce ambiguity and reduced execution efficiency compared with structured communication.Source: §2.2.2 · p. 8

2.3 The Vision of Lifelong, Self-Evolving Agentic Systems

• Existing LLM and multi-agent systems still rely on handcrafted workflows, fixed protocols, and curated toolchains, constraining adaptation as tasks, resources, and goals change.This motivates lifelong self-evolving systems that reduce manual configuration in dynamic, open-ended environments.Source: §2.3 · p. 9
• Multi-Agent Self-Evolving systems close the deployment-improvement loop by autonomously refining prompts, memory, tool-use strategies, and interaction topologies in response to environmental feedback and meta-rewards.They monitor performance and safety, preserve or enhance capabilities through controlled updates, and adapt over their lifetimes to shifting tasks, environments, and resources.Source: §2.3 · p. 9

3 A Conceptual Framework of MASE

• The framework provides a generalisable abstraction of agent evolution and optimisation methods, supporting comprehensive understanding and comparative analysis across approaches.Source: §3 · p. 9

3.1 Overview of the Self-Evolving Process

• Self-evolution forms a closed iterative loop in which an agent system receives environmental feedback, is updated by an optimiser, and is redeployed until performance or convergence criteria are met.Updates may adjust LLM parameters, prompts, system structure, or synthesised training examples; EvoAgentX automates this process (Wang et al., 2025i).Source: §3.1 · pp. 9–10, Figure 3 · p. 10

3.2 System Inputs

• System inputs define the optimisation problem through task requirements, constraints, and data, supporting task-level optimisation with T and Dtrain or instance-level optimisation with (x, y, C).When labelled task data are unavailable, LLM-generated surrogate examples can support optimisation (Huang et al., 2025; Zhao et al., 2025a; Liu et al., 2025b).Source: §3.2 · pp. 10–11

3.3 Agent Systems

• The agent system A may be single-agent or multi-agent and can be optimised through its LLM, prompts, memory, tools, or jointly tuned components and inter-agent topology.Examples include LLM finetuning, prompt and tool selection, joint LLM-prompt optimisation, and joint prompt-topology optimisation (Zelikman et al., 2022; Tong et al., 2024; Lai et al., 2024b; Yang et al., 2024a; Yuan et al., 2025b; Soylu et al., 2024; Zhang et al., 2025j; Zhou et al., 2025a).Source: §3.3 · p. 11

3.4 Environments

• The environment supplies the operating context, receives agent actions, and produces feedback from task-specific metrics or LLM-based proxy evaluation when labels or ground truth are unavailable.Environments range from benchmark datasets to dynamic real-world settings, including code execution systems, literature databases, simulations, and laboratory equipment (Liu et al., 2023a; Yehudai et al., 2025).Source: §3.4 · p. 11

3.5 Optimisers

• An optimiser is defined by a search space S and optimisation algorithm H, whose combination determines how candidate agent configurations are explored and adapted toward better performance.Search spaces may cover prompts, tools, parameters, or architectures, while algorithms include heuristics, gradient descent, Bayesian optimisation, MCTS, reinforcement learning, evolutionary strategies, and learning-based policies; designs differ across single-agent, multi-agent, and domain-specific settings.Source: §3.5 · pp. 11–12

• Single-agent optimisation targets an individual agent system, with the central challenge of designing optimisers that select components, capabilities, and effective update strategies.The relevant design dimensions are the optimisation search space, the capabilities to enhance, and the optimisation algorithm.Source: §4 · p. 12
• The section organises approaches into four categories—LLM Behaviour, Prompt, Memory, and Tool optimisation—based on the targeted component and resulting search space.Figure 4 presents this component-based overview of single-agent optimisation approaches.Source: §4 · p. 12, Figure 4 · p. 12
• LLM Behaviour optimisation improves reasoning and planning through parameter tuning or test-time scaling, while Prompt optimisation adapts prompts for more accurate, task-relevant outputs.Source: §4 · p. 12
• Memory optimisation enhances storing, retrieving, and reasoning over historical information or external knowledge, whereas Tool optimisation improves tool use or enables new tool creation and configuration.Source: §4 · p. 12

4.1.1 Training-Based Behaviour Optimisation

• Training-based optimisation targets the gap between LLM fluency and complex reasoning by teaching models to evaluate and refine responses for multi-step inference and decision-making.It directly updates model parameters using supervised fine-tuning or reinforcement learning.Source: §4.1.1 · p. 14, §4.1 · p. 14
• SFT trains on annotated reasoning trajectories from self-generated rollouts or stronger teachers, while RL optimises sequential reasoning using preference, verifiable, proof-assistant, or self-generated rewards.Representative systems include STaR (Zelikman et al., 2022), NExT (Ni et al., 2024), Deepseek-Prover (Xin et al., 2024), Agent Q (Putta et al., 2024), DeepSeek-R1 (Guo et al., 2025), Absolute Zero (Zhao et al., 2025a), and R-Zero (Huang et al., 2025).Source: §4.1.1 · p. 14

4.1.2 Test-Time Behaviour Optimisation

• Test-time compute lets models refine reasoning during inference without additional training, addressing constrained training resources and non-fine-tunable API models by enabling them to think longer.Its two strategies are external-feedback guidance and multi-candidate generation followed by verifier-based selection, which are closely related because feedback can act as a verifier.Source: §4.1.2 · p. 15
• Feedback-based inference uses outcome- or step-level verification: tools validate code or proofs, learned verifiers rank outputs, and process reward models detect intermediate errors that outcome feedback can miss.Step-level feedback addresses unfaithful reasoning, where incorrect chains produce correct final answers, and generally improves more than outcome-only feedback.Source: §4.1.2 · p. 15
• Search-based inference explores multiple reasoning trajectories using voting, beam search, trees, graphs, forests, or dynamic memory, with critic models balancing exploration and exploitation.Examples include CoT-SC (Wang et al., 2023b), DBS, Tree-of-Thoughts (Yao et al., 2023a), Forest-of-Thought (Bi et al., 2025), Graph-of-Thoughts (Besta et al., 2024), and Buffer-of-Thoughts (Yang et al., 2024c).Source: §4.1.2 · p. 15

4.2 Prompt Optimisation

• Prompt sensitivity makes robust, generalisable agent design difficult, motivating automatic prompt search across four optimisation categories.The categories are edit-based, generative, text gradient-based, and evolutionary methods.Source: §4.2 · pp. 15–16

4.2.1 Edit-Based Prompt Optimisation

• Edit-based methods locally refine human-written prompts through insertion, deletion, substitution, or phrase-level operations while preserving core instruction semantics.GRIPS uses phrase edits, Plum adds simulated annealing, mutation, and crossover, and TEMPERA learns editing policies with reinforcement learning.Source: §4.2.1 · p. 16

4.2.2 Generative Prompt Optimisation

• Generative methods use LLMs to create entirely new prompts, exploring broader and more diverse regions of prompt space than local editing.They condition generation on base prompts and optimisation signals, including rewriting rules, examples, descriptions, scores, objectives, and success or failure cases.Source: §4.2.2 · p. 16
• Generative optimisation combines guidance signals with Gibbs sampling, Monte Carlo tree search, Bayesian optimisation, or neural bandits to explore prompt space efficiently.PromptAgent uses MCTS for strategic planning, while MIPRO applies Bayesian optimisation to instruction candidates and few-shot demonstrations.Source: §4.2.2 · p. 16
• Beyond frozen LLM generation, reinforcement-learning approaches train policy models to iteratively generate or refine prompts from previous outcomes.Retroformer summarises root causes of failed cases to guide prompt refinement.Source: §4.2.2 · p. 16

4.2.3 Text Gradient-Based Prompt Optimisation

• Text gradient methods replace numerical parameter gradients with natural-language feedback that directs prompt or broader symbolic-system updates.ProTeGi critiques prompts and performs opposite-direction edits with beam search and bandit selection; TextGrad extends textual feedback to prompts, code, and symbolic variables, while agent symbolic learning optimises prompts, tools, and workflows.Source: §4.2.3 · pp. 16–17

4.2.4 Evolutionary Prompt Optimisation

• Evolutionary methods maintain populations of candidate prompts and apply mutation, crossover, and selection to evolve high-performing instructions.EvoPrompt adapts genetic algorithms and differential evolution, while Promptbreeder uses predefined or LLM-generated mutation prompts to guide task-prompt changes.Source: §4.2.4 · p. 17

4.3 Memory Optimisation

• The survey frames inference-time memory optimisation as dynamic control over memory scheduling, updating, reuse, and retention, addressing context-window limits, forgetting, context drift, and hallucination.Unlike training-time fine-tuning or knowledge editing, these approaches optimise memory utilisation without modifying model parameters.Source: §4.3 · p. 17

4.3.1 Short-term Memory Optimisation

• Short-term methods compress, summarise, selectively retain, or retrieve active-context information to improve local coherence and context efficiency during extended interactions.Representative mechanisms include recursive summarisation, dialogue-level memory, compressed traces, structured retrieval, hierarchical recency-based updates, and episodic reflection, but short-term memory alone cannot preserve knowledge across sessions.Source: §4.3.1 · pp. 17–18

4.3.2 Long-term Memory Optimisation

• Long-term memory systems provide persistent storage and retrieval across sessions through RAG, agentic retrieval, trajectory memory, vector or graph structures, symbolic databases, and self-evolving memory architectures.Their optimisation also depends on scalable indexing, pruning, compression, and inference-time policies that decide what, when, and how to store, update, or discard; examples include reinforcement-learning consolidation, autonomous memory organisation, and collaborative knowledge transfer.Source: §4.3.2 · pp. 18–19

4.4 Tool Optimisation

• Research divides tool optimisation into improving agent–tool interaction and modifying or creating tools that better match task requirements.Agent interaction is addressed through training, prompting, and reasoning algorithms, while tool-centred methods adapt the available toolset rather than adapting tasks to fixed tools.Source: §4.4 · p. 19, §4.4.3 · p. 21

4.4.1 Training-Based Tool Optimisation

• Training-based methods teach LLMs tool use through supervised trajectories or reinforcement learning, improving planning, execution, and generalisation across tools and task configurations.SFT uses trajectories containing queries, reasoning, tool calls, and answers; RL addresses limited data coverage through interactive feedback and includes multi-tool, multimodal, and long-horizon strategies.Source: §4.4.1 · pp. 19–20

4.4.2 Inference-Time Tool Optimisation

• Inference-time optimisation improves tool use without changing LLM parameters by refining tool documentation and prompts or guiding test-time search and planning.Methods unify or iteratively refine documentation, jointly optimise instructions and descriptions, prune low-value search branches, reorganise tools into toolkits, and request missing capabilities on demand.Source: §4.4.2 · pp. 20–21

4.4.3 Tool Functionality Optimisation

• Tool functionality optimisation expands agents’ action spaces by generating, modifying, or continually refining executable tools and their documentation for unseen or evolving tasks.Approaches reuse prior code, treat tools as learnable components, create Multi-Component Programs, and adapt visual tools through inference, reflection, learning, and human feedback.Source: §4.4.3 · p. 21

5 Multi-Agent Optimisation

• Self-evolving systems replace fixed multi-agent architectures with feedback-driven search over agent prompts, topologies, collaborative strategies, and LLM capabilities, while balancing accuracy, efficiency, and safety.The section frames workflow design as optimisation across structural, semantic, and capability spaces, spanning methods from evolutionary algorithms to reinforcement learning.Source: §5 · pp. 21–22, §5.2 · p. 23

5.2 Self-Evolving Multi-Agent System

• Overall, the field expands the searchable object from instructions and communication protocols to adaptive workflow architectures and the underlying models’ cooperative capabilities.This progression reframes optimality as a dynamic, multi-objective question rather than merely finding a single best configuration.Source: §5 · pp. 21–22, §5.2 · p. 23, §5.2.4 · p. 25

5.1 Manually Designed Multi-Agent Systems

• Handcrafted workflows use parallel, hierarchical, or debate-based collaboration, but fixed structures and high coordination costs motivate adaptive alternatives; single large LLMs can sometimes match complex multi-agent systems.Parallel workflows can scale through asynchronous execution, hierarchical systems suit context-dependent subtasks, and confidence-gated debate reduces inference costs.Source: §5.1 · pp. 22–23

5.2.1 Multi-Agent Prompt Optimisation

• Prompt optimisation refines role instructions and intermediate outputs, extending single-agent methods to coordinated teams through assertion-driven feedback, backtracking, and structured meta-agent dialogue.DSPy (Singhvi et al., 2023) uses assertion-guided self-correction, while AutoAgents (Chen et al., 2024b) optimises specialised roles and execution plans.Source: §5.2.1 · p. 23

5.2.2 Topology Optimisation

• Topology optimisation treats communication structure as a first-class search target, covering executable code workflows and communication graphs that can be synthesised, adapted, pruned, or secured.Methods use reinforcement learning, Monte Carlo Tree Search, evolutionary operators, graph generation, online restructuring, and pruning to trade off performance, adaptability, token cost, and safety.Source: §5.2.2 · pp. 23–24

5.2.3 Unified Optimisation

• Unified optimisation jointly evolves prompts and topologies through code representations, evolutionary search, structured alternating procedures, debate, inference-time refinement, and learned supernets or layered architectures.These approaches differ in adaptation granularity and trade-offs among flexibility, search efficiency, performance, and cost.Source: §5.2.3 · pp. 24–25

5.2.4 LLM Backbone Optimisation

• LLM-backbone optimisation trains reasoning and cooperation through debate trajectories, self-play, supervised fine-tuning, preference optimisation, and reinforcement learning with task-specific communication rewards.Sirius and MALT use different treatments of failed trajectories, while MaPoRL, MARFT, and MARTI explicitly optimise cooperative structures or reward functions.Source: §5.2.4 · pp. 25–26
• Collaboration-oriented training targets communication quality and efficiency because out-of-the-box LLMs may lack effective teamwork skills.OPTIMA (Chen et al., 2025h) reports a 2.8× performance gain with less than 10% of the token cost on information-intensive tasks, while COPPER and MaPoRL use learned reflection or reinforcement signals.Source: §5.2.4 · p. 26

6 Domain-Specific Optimisation

• Domain-specific agent systems require tailored optimisation because specialised tasks, knowledge bases, data modalities, and operational constraints shape how agents are designed and evolved.The survey covers biomedicine (Almansoori et al., 2025), programming (Tang et al., 2024), scientific research (Pu et al., 2025), game-playing (Belle et al., 2025), computer use (Sun et al., 2025), and finance and legal research.Source: §6 · p. 26

6.1 Domain-Specific Optimisation in Biomedicine

• Biomedical agent optimisation aligns agent behaviour with real-world clinical requirements, especially in medical diagnosis and molecular discovery.Representative studies include medical diagnosis (Almansoori et al., 2025; Zhuang et al., 2025) and molecular discovery (M. Bran et al., 2024; Inoue et al., 2025).Source: §6.1 · p. 26

6.1.1 Medical Diagnosis

• Medical diagnosis agents handle incomplete or ambiguous information through multi-turn questioning, diagnostic hypothesis generation, and integration of external medical knowledge or specialised tools.These systems support autonomous diagnostic dialogues and evidence-based reasoning under uncertain clinical conditions (Li et al., 2024c; Chen et al., 2025i; Zuo et al., 2025; Ghezloo et al., 2025).Source: §6.1.1 · pp. 26–27
• Multi-agent diagnostic systems use simulation-driven or collaborative designs to model clinical workflows, assign specialised roles, integrate suggestions, and support reflective team-based reasoning.MedAgentSim (Almansoori et al., 2025) combines experience replay, chain-of-thought ensembling, and CLIP-based semantic memory; PathFinder (Ghezloo et al., 2025), MDAgents (Kim et al., 2024), and MDTeamGPT (Chen et al., 2025c) provide complementary workflow and collaboration mechanisms.Source: §6.1.1 · p. 27
• Diagnostic optimisation also combines multimodal tool integration, criteria-guided planning, and feedback-adaptive reasoning to improve specialisation, clinical reliability, and interactive reasoning.MMedAgent (Li et al., 2024a) dynamically incorporates specialised tools, MedAgent-Pro (Wang et al., 2025l) integrates multimodal evidence, and Zhuang et al. (2025) adjusts reasoning using diagnostic-result feedback.Source: §6.1.1 · p. 27

6.1.2 Molecular Discovery and Symbolic Reasoning

• Molecular discovery agents use chemical tools, structured memory, and multi-agent collaboration to perform symbolic reasoning over chemical structures, reactions, and pharmacological constraints.CACTUS (McNaughton et al., 2024) uses RDKit for chemically valid outputs, while LLM-RDF (M. Bran et al., 2024) coordinates specialised agents for literature mining, synthesis planning, and reaction optimisation.Source: §6.1.2 · p. 27
• Memory-enabled and coordinated molecular agents learn from prior solutions, convert failures into structured updates, and combine predictors, knowledge graphs, literature search, and modular evaluation workflows.ChemAgent, OSDA Agent, DrugAgent, and LIDDIA (Averly et al., 2025) support retrieval, self-reflection, interpretable reasoning, and multi-objective molecule evaluation.Source: §6.1.2 · pp. 27–28

6.2 Domain-Specific Optimisation in Programming

• Programming optimisation targets code refinement and debugging by aligning agent behaviours with established software-engineering workflows.Representative studies cover code refinement (Rasheed et al., 2024; Tang et al., 2024; Pan et al., 2025b) and debugging (Lee et al., 2024a; Adnan et al., 2025).Source: §6.2 · p. 28

6.2.1 Code Refinement

• Code-refinement agents iteratively improve quality while preserving functionality through self-critique, memory-enabled collaboration, tool integration, and interactive reasoning.Self-Refine (Madaan et al., 2023), LLM-Surgeon (van der Ouderaa et al., 2023), AgentCoder (Huang et al., 2023a), CodeAgent (Tang et al., 2024), CodeCoR (Pan et al., 2025b), OpenHands (Wang et al., 2025g), and VFlow (Wei et al., 2025b) exemplify these strategies.Source: §6.2.1 · p. 28

6.2.2 Code Debugging

• Debugging agents use execution traces and modular role specialisation to localise faults, reason about failures, validate semantics, and revise code autonomously.Self-Debugging (Chen et al., 2024c) and Self-Edit (Zhang et al., 2023a) use runtime feedback, while PyCapsule (Adnan et al., 2025), Self-Collaboration (Dong et al., 2024), RGD (Jin et al., 2024), and FixAgent (Lee et al., 2024a) structure specialised or hierarchical workflows.Source: §6.2.2 · pp. 28–29

6.3.1 Financial Decision-Making

• Financial agents integrate conceptual collaboration, specialised roles, external tools, and real-time data to support decision-making under volatile conditions (Yu et al., 2024b; Wang et al., 2024j; Yang et al., 2024b).FinCon uses verbal reinforcement and domain-adaptive fine-tuning, PEER combines expert, retriever, and controller roles, and FinRobot connects strategies with executable financial models.Source: §6.3.1 · p. 29
• Financial reporting systems combine sentiment modules, rule-based validation, and iterative retrieval, validation, and synthesis to improve reliable and interpretable decision support (Xing, 2025; Tian et al., 2025).These designs address heterogeneous information and domain-specific reporting guidelines through real-world feedback.Source: §6.3.1 · p. 29

6.3.2 Legal Reasoning

• Legal collaborative frameworks assign specialised roles and simulate adversarial courtroom procedures, while AgentCourt improves debate quality and procedural realism through reflective self-play (Sun et al., 2024b; Di Martino et al., 2023; Chen et al., 2025a).LawLuo uses a central controller for procedural consistency and legal correctness; Multi-Agent Justice Simulation and AgentCourt model role-based courtroom interactions.Source: §6.3.2 · pp. 29–30
• LegalGPT and AgentsCourt ground collaborative reasoning in interpretable chains of thought, codified rules, and case precedents to support transparent and legally reliable decisions (He et al., 2024).These systems combine structured legal reasoning with domain knowledge augmentation and courtroom debate simulation.Source: §6.3.2 · p. 30

7 Evaluation

• Agent evaluation increasingly combines standardised task benchmarks, safety audits, and LLM-as-a-judge methods, with fine-grained signals guiding optimisation, prompt refinement, and dataset augmentation.Evaluation therefore functions as a dynamic feedback mechanism rather than only a static endpoint.Source: §7 · p. 30

7.1.1 Tool and API-Driven Agents

• Tool-agent benchmarks assess API-call correctness and efficiency, while newer multi-step environments evaluate planning, coordination, and decision-making processes; overfitting to unseen tool schemas remains a challenge.ToolBench, API-Bank, MetaTool, ToolQA, GTA, and AppWorld exemplify this progression.Source: §7.1.1 · p. 30

7.1.2 Web Navigation and Browsing Agents

• Web benchmarks increasingly test navigation, interface adaptability, and textual-visual integration in simulated and live environments, but dynamic websites still hinder reproducibility and generalisation.Intermediate measures such as sub-goal completion and robustness assessments complement final task outcomes.Source: §7.1.2 · p. 30

7.1.3 Multi-Agent Collaboration and Generalists

• Multi-agent and generalist benchmarks evaluate collaboration, competition, decentralised coordination, and cross-domain adaptability, while exposing metric aggregation, narrow-scenario overfitting, and leaderboard-unification challenges.Efficient Agents further examines the efficiency–effectiveness trade-off while reducing operational costs.Source: §7.1.3 · pp. 30–31

7.1.4 GUI and Multimodal Environment Agents

• GUI and multimodal benchmarks require complex actions combining language, visual perception, and API use, but interface variability makes standardisation difficult and leaves agents brittle.Evaluations measure task success, state management, perception accuracy, and adaptability to GUI changes.Source: §7.1.4 · p. 31

7.1.5 Domain-Specific Task Agents

• Domain-specific benchmarks evaluate planning, tool use, specialised competencies, policy adherence, and domain constraints, yet inconsistent metrics and generalisation gaps persist across coding, science, enterprise, and clinical tasks.Examples include SWE-bench, DataSciBench, MLGym, WorkBench, OpenAGI, SUPER, AgentClinic, and MMedAgent.Source: §7.1.5 · p. 31

7.2.1 LLM-as-a-Judge

• LLM judges provide scalable evaluation through pointwise scoring or pairwise comparison and can approach human-judgment agreement, but prompt sensitivity, bias, and shallow output focus limit reliability.Multi-agent deliberation and structured meta-evaluation are proposed to improve calibration and capture multi-step reasoning.Source: §7.2.1 · p. 31

7.2.2 Agent-as-a-Judge

• Agent-as-a-Judge evaluates complete reasoning trajectories, intermediate actions, and artefacts, aligning more closely with expert judgments and reducing review time and cost, but deployment remains complex beyond code generation.Specialised modules construct reasoning graphs and validate hierarchical requirements in DevAI evaluations.Source: §7.2.2 · pp. 31–32

8.1.1 Endure – Safety Adaptation

• Self-evolving agents require evolution-aware safety mechanisms because metric-focused optimisation can cause unintended behaviour, privacy breaches, misalignment, and instability beyond static legal frameworks.Proposed safeguards include adaptive licences, provable-safety sandboxes, audits, and legal protocols tracking self-directed evolution.Source: §8.1.1 · pp. 32–33

8.1.2 Excel – Performance Preservation

• Performance preservation remains difficult because domain-specific ground truth may be disputed, multi-agent optimisation is costly and unstable, and prompts or topologies transfer poorly across LLM backbones.These limitations complicate trustworthy feedback, efficiency–effectiveness trade-offs, scalability, and production reuse.Source: §8.1.2 · p. 33

8.1.3 Evolve – Autonomous Optimisation

• Autonomous optimisation must extend beyond text-only methods and fixed toolsets to support multimodal, spatial, and continuous environments, including internal world models and co-evolving tools.Agents need capabilities for autonomous tool discovery, adaptation, and creation alongside perceptual–temporal reasoning.Source: §8.1.3 · p. 33

8.2 Future Directions

• The proposed research agenda links future directions to the MOP→MOA→MAO→MASE paradigm shift and targets fully autonomous self-evolution through open-ended simulated environments.Agents would interact iteratively, receive feedback, and refine prompts, memory, tools, and workflows via closed-loop optimisation.Source: §8.2 · p. 33
• Future systems should advance tool use from static invocation toward adaptive selection, composition, and creation using reinforcement learning, feedback, and robust evaluation.This direction is explicitly associated with the MAO→MASE transition.Source: §8.2 · p. 33
• Future benchmarks should reflect real-world complexity through interaction-based and longitudinal assessment aligned with long-term improvement signals, while MAS optimisation jointly models performance and resource constraints.The goal is deployment under strict latency, cost, or energy budgets.Source: §8.2 · p. 33
• Domain-aware evolution should incorporate heterogeneous knowledge, bespoke evaluation, and regulatory compliance for scientific, medical, legal, and educational applications.This direction is associated with MASE and addresses domain-specific constraints.Source: §8.2 · p. 33
• Together, these directions call for high-performing, domain-adaptive, safe, regulation-aware, and self-sustaining optimisation pipelines grounded in the Three Laws across the MOP→MOA→MAO→MASE trajectory.The intended outcome is lifelong autonomous agents that endure, excel, and evolve throughout their operational lifetimes.Source: §8.2 · p. 33

• The survey positions self-evolving agents along a four-stage trajectory from Model Offline Pretraining (MOP) and Model Online Adaptation (MOA) through Multi-Agent Orchestration (MAO) to Multi-Agent Self-Evolving (MASE).This trajectory captures the shift from static, human-configured models toward dynamic, autonomous ecosystems.Source: §9 · p. 34
• Its unified framework models agent evolution as a feedback loop among Inputs, Agent System, Objectives, and Optimisers, supporting systematic analysis of optimisation techniques, domain strategies, and evaluation methods.These components determine how agents improve through continual interaction with their environments.Source: §9 · p. 34
• The proposed Three Laws—Endure, Excel, and Evolve—make safety adaptation, performance preservation, and autonomous evolution practical design constraints for the MASE paradigm.They guide self-improvement toward safety, effectiveness, alignment, and continuous open-ended adaptation.Source: §9 · p. 34
• Realising lifelong agents in dynamic environments will require scalable optimisation, lifelong evaluation, safe coordination across heterogeneous agents, and adaptation mechanisms for unforeseen domains.These needs span applications including scientific discovery, software engineering, and human–AI collaboration.Source: §9 · p. 34
• The survey calls for an ecosystem of self-evolving agents that move beyond task execution toward autonomous, resilient, trustworthy systems that live, learn, and last.It links this goal to aligning technical innovation with principled self-evolution.Source: §9 · p. 34