AgentFactory: A Self-Evolving Framework Through Executable Subagent Accumulation and Reuse

AgentFactory: A Self-Evolving Framework Through Executable Subagent Accumulation and Reuse Zhang Zhang 1,3 , Shuqi Lu 3 , Hongjin Qian 1,3 , Di He 1 , Zheng Liu 2,3 1 Peking University, 2 Hong Kong Polytechnic University 3 Beijing Academy of Artificial Intelligence zzhang@stu.pku.edu.cn, zhengliu1026@gmail.com Abstract Building LLM-based agents has become in- creasingly important. Recent works on LLM- based agent self-evolution primarily record suc- cessful experiences as textual prompts or re- flections, which cannot reliably guarantee ef- ficient task re-execution in complex scenarios. We propose AgentFactory, a new self-evolution paradigm that preserves successful task solu- tions as executable subagent code rather than textual experience. Crucially, these subagents are continuously refined based on execution feedback, becoming increasingly robust and ef- ficient as more tasks are encountered. Saved subagents are pure Python code with standard- ized documentation, enabling portability across any Python-capable system. We demonstrate that AgentFactory enables continuous capabil- ity accumulation: its library of executable sub- agents grows and improves over time, progres- sively reducing the effort required for simi- lar tasks without manual intervention. Our implementation is open-sourced athttps: //github.com/zzatpku/AgentFactory, and our demonstration video is available at https://youtu.be/iKSsuAXJHW0. 1 Introduction Large Language Models (LLMs) have demon- strated remarkable capabilities in reasoning, plan- ning, and problem-solving (Anthropic, 2026; Ope- nAI, 2026; Gemini team, 2026). The development of LLM-based agents that can interact with exter- nal tools and environments has become a critical research area (Yao et al., 2023; Wang et al., 2024a; Xi et al., 2025). These agents must not only rea- son about tasks but also execute actions, retrieve information, and adapt to dynamic environments. Existing frameworks for building LLM agents, such as LangChain (Chase, 2022) and AutoGPT (Significant Gravitas, 2023), provide valuable ab- stractions for connecting LLMs with external tools. However, these frameworks treat agent behavior as static—knowledge gained during execution is not preserved for future use. To address this, re- cent works have explored self-evolving agents that record and leverage past experience through verbal reflection, iterative output refinement, or reasoning bootstrapping (Shinn et al., 2023; Madaan et al., 2023; Zelikman et al., 2022). However, these ap- proaches primarily record successful experiences as textual prompts, verbal reflections, or reason- ing traces. For complex real-world tasks, such textual experience cannot reliably guarantee ef- ficient task re-execution. Recent advances have begun exploring code-based self-evolution as an alternative. AlphaEvolve (Novikov et al., 2025) demonstrates the power of code-based evolution for high-complexity scientific discovery, and the Darwin Gödel Machine (Zhang et al., 2025b) ex- plores open-ended recursive self-improvement of agent internals. However, these methods are pri- marily tailored for solving highly specialized sci- entific or meta-reasoning problems. Nonetheless, a vast array of general-purpose workflows still relies on routine, procedural tasks that demand a different approach to evolution. We observe that although daily user tasks appear diverse and complex, most can be decomposed into a set of reusable subtasks—for instance, schedul- ing a meeting, conducting a literature survey, or file manipulations. This observation motivates us to propose a self-evolution approach centered on creating and accumulating subagents: we decom- pose tasks into subtasks and construct specialized subagents for each one. Successfully executed sub- agents are saved as executable code, forming a growing library of reusable capabilities. Crucially, these subagents are autonomously refined based on execution feedback from subsequent tasks, making them increasingly robust and general-purpose over time. Furthermore, since all saved subagents are pure Python code with standardized documenta- tion, they can be exported and deployed across any 1 arXiv:2603.18000v1 [cs.AI] 18 Mar 2026 Python-capable system, enabling cross-platform capability transfer. Based on this design, we propose AgentFactory, a self-evolving framework that implements a three- phase lifecycle: (1) Install: Construct subagents from scratch to solve initial problems, building a library of reusable subagents. (2) Self-Evolve: When encountering similar tasks, detect limita- tions in saved subagents and autonomously im- prove them to be more general-purpose and robust. (3) Deploy: Export mature subagents as standalone Python modules for use in other AI frameworks. We observe that after executing a small number of initial tasks, the system begins to accumulate a collection of subagents that can be reused to assist with many common real-world tasks. In this way, solving a few initial tasks provides a practical start- ing point: the subagents constructed during these early runs can be leveraged to support subsequent requests without additional manual setup. This au- tomatically running pipeline creates an efficient self-evolution system that becomes increasingly helpful for everyday user needs. Our contributions are as follows: 1. We propose a three-phase self-evolving pipeline (Install→Self-Evolve→Deploy) that systematically builds, improves, and ex- ports functional agents. Our framework en- ables the system to accumulate and improve complex task-solving capabilities rather than merely recording procedural experiences. 2.We implement AgentFactory and demonstrate on diverse tasks that the system becomes in- creasingly effective: saved subagents are di- rectly reused to solve new problems (faster problem-solving), and they are iteratively im- proved based on execution feedback (growing capabilities). 2 Related Work Our work relates to multi-agent systems, self- evolving mechanisms, and skill accumulation. Multi-Agent Systems Multi-agent frameworks such as AutoGen (Wu et al., 2024), MetaGPT (Hong et al., 2024), and ChatDev (Qian et al., 2024) enable collaboration among specialized agents through predefined workflows. Recent research has shifted toward dynamic orchestration and topologi- cal optimization. For instance, AgentVerse (Chen et al., 2024) mimics human group dynamics for expert recruitment, while DyLAN (Liu et al., 2023) introduces an unsupervised metric for dynamic agent team optimization. Furthermore, GPTSwarm (Zhuge et al., 2024) treats agents as optimizable graphs, and frameworks like CrewAI and Lang- Graph support role-based task execution and cyclic state management. Self-Evolution and Skill AccumulationRecent advancements enable agents to continuously im- prove through self-evolution and persistent mem- ory. Evolutionary approaches optimize specific agent components, including prompts (Wang et al., 2024b; Guo et al., 2024; Fernando et al., 2024), reasoning strategies (Shinn et al., 2023; Zelikman et al., 2022), architectures (Hu et al., 2024; Zhang et al., 2025c; Li et al., 2024), and algorithmic code (Novikov et al., 2025; Zhang et al., 2025b). Con- currently, skill accumulation methods maintain ex- perience through structured memory mechanisms (Tang et al., 2025; Zhou et al., 2025; Xu et al., 2025) or by saving executable tool-level skills, as demonstrated by Voyager (Wang et al., 2023). 3 AgentFactory: Method As introduced earlier, AgentFactory follows a three-phase lifecycle—Install, Self-Evolve, and De- ploy—that governs how agents are constructed, im- proved, and exported. In this section, we describe the concrete implementation of each component and how they support the three phases of this life- cycle. 3.1 Architecture Overview AgentFactory implements a self-evolving frame- work built upon a unified skill that aligns with the emerging Agent Skills open standard (Zhang et al., 2025a), consisting of three main components: the Meta-Agent orchestrator, the Skill System, and the Workspace Manager. The Meta-Agent serves as the central orches- trator that decomposes complex problems into sub-problems and assigns them to specialized sub- agents. Crucially, when creating a subagent, the Meta-Agent dynamically selects and allocates the relevant tools from the skill library to that subagent, rather than exposing the entire tool set and letting the subagent discover tools on its own. This design reduces the subagent’s search space and ensures that each subagent operates with a focused, task- appropriate toolkit. The Meta-Agent maintains a 2 Q1 Subagent Pool Q2 Other System Export Evolve AgentFactoryAgentFactory Save Save Create Solve Question Search Figure 1: Overview of the AgentFactory pipeline. The figure illustrates two boundary cases. Q1 represents a task where no relevant subagents exist: the Meta-Agent creates new subagents from scratch and saves them to the subagent pool. Q2 represents a task where a matching subagent already exists: the Meta-Agent reuses it and modifies it through the self-evolving process to handle the new requirements. In practice, most tasks fall between these two extremes—the Meta-Agent reuses some existing subagents for part of the problem while constructing new ones for the remainder. Over multiple rounds, each task may create or modify several subagents, gradually building a rich and capable subagent pool that can eventually be exported for use by other systems. execution history and can iteratively refine the sub- agents based on execution results. The Skill System unifies all operations as skills with three levels: Meta Skills Operations for agent orches- trationincludingcreate_subagent,get- _skill_description,run_subagent,modify- _subagent,finish,list_saved_subagents, andview_subagent_code. These built-in skills remain fixed and provide fundamental primitives for agent lifecycle management. Tool Skills Built-in tools includingweb- _search(Serper),web_reading(Jina),browser- _automation(Playwright), andshell_command. Like meta skills, these are static and serve as the primitive building blocks from which subagents are constructed. All tool skills include detailed documentation on parameters and return formats. Subagent SkillsReusable modules dynamically created and refined during execution for problem- solving. Unlike the two fixed skill categories above, subagent skills are generated as executable Python scripts that encapsulate successful task-solving pat- terns, and they evolve over time—new subagents are added as the system encounters new tasks, and existing ones are modified based on execution feed- back to become more robust and general-purpose. The Workspace Manager provides isolated exe- cution environments for each task. Each task exe- cution operates in a dedicated workspace directory, ensuring that subagents can safely create, modify, and execute code without interfering with one an- other or corrupting the shared skill library. This iso- lation is essential for reliable self-evolution: when a subagent is being modified or tested, any failures are contained within the workspace and do not af- fect the saved skills or other concurrently running tasks. Upon successful completion, results and im- proved subagents are promoted from the workspace to the persistent skill library. 3.2 Phase 1: Install — Constructing Subagents from Scratch When AgentFactory encounters a new problem that cannot be solved by existing skills, it enters the Install phase. The Meta-Agent analyzes the task requirements, decomposes it into sub-problems, and dynamically constructs specialized subagents to address each sub-problem. The installation process works as follows: Task Analysis The Meta-Agent analyzes the user’s query and identifies the capabilities required to solve the task. Subagent Construction The Meta-Agent de- composes the task into multiple sub-problems and invokescreate_subagentfor each one, dynam- ically generating a specialized Python script that encapsulates the reasoning logic and tool invoca- tions needed to address that sub-problem. A single task typically results in the creation of several sub- agents, each responsible for a distinct capability. 3 ExecutionThe newly created subagent executes its assigned task using available tool skills. Persistence Upon successful completion, the Meta-Agent evaluates the subagent and decides which skills to save as reusable skills. Each saved subagent consists of pure Python code with an ac- companying SKILL.md file documenting its func- tionality, parameters, and usage. At this phase, AgentFactory functions as an "agent factory"—it builds functional agents from scratch rather than relying on pre-defined tem- plates. The created subagents are immediately usable within the system and become part of the growing skill library. This mechanism fundamentally differs from experience recording approaches like Reflexion (Shinn et al., 2023). Rather than simply maintain- ing verbal reflections or histories, AgentFactory actively transforms successful execution patterns into reusable code that can be immediately applied to new problems. 3.3 Phase 2: Self-Evolve — Improving Subagents Through Feedback As AgentFactory accumulates more saved sub- agents, it enters the Self-Evolve phase when en- countering tasks similar to previously solved ones. Instead of building from scratch, the Meta-Agent retrieves relevant saved subagents and attempts to reuse them. When a saved subagent fails to handle a new vari- ation of a task or produces suboptimal results, the Meta-Agent performs autonomous improvement: Retrieval The Meta-Agent uses the skill list_saved_subagentsto discover previously saved skills that may be relevant to the current task. AssessmentThe Meta-Agent runs the candidate subagent and evaluates its performance against the new requirements. Feedback Analysis After running the subagent, if the subagent fails, the Meta-Agent analyzes the execution feedback to identify specific failure modes. Autonomous Modification The Meta-Agent in- vokesmodify_subagentto refine the subagent code, making it more robust. This may involve adding error handling, extending functionality to handle edge cases, or restructuring the logic to be more adaptable. Validation The modified subagent is tested against the current task to verify the improvement. This self-evolution process enables AgentFac- tory to continuously improve its capabilities with- out external intervention. Each iteration makes the subagent more general and robust, effectively "in- stalling" improved versions of the agent into the skill library. The self-evolving mechanism extends the foun- dational "generate-feedback-modify" loop estab- lished by Self-Refine (Madaan et al., 2023) from single-output refinement to agent-level improve- ment. Unlike prior approaches that focus on ver- bal reflection or prompt optimization (Yang et al., 2023; Zelikman et al., 2022), AgentFactory modi- fies actual executable agent code, leading to tangi- ble capability improvements. 3.4 Phase 3: Deploy — Exporting Subagents for External Use The final phase is Deploy, where mature and val- idated subagents are exported for use in other AI frameworks. Since all saved subagents are pure Python code with accompanying SKILL.md docu- mentation, they can be directly imported and used by any system capable of executing Python. Specif- ically, AgentFactory supports the following deploy- ment scenarios: Standalone ExecutionEach subagent can be run independently as a Python script without depending on the AgentFactory runtime. Framework Integration Saved subagents can be integrated into other agent frameworks (e.g., LangChain (Chase, 2022), AutoGen (Wu et al., 2024), or Claude Code 1 . Integration is achieved by providing the external agent with prompts that explain how to invoke the subagent scripts and how to consult the SKILL.md documentation to under- stand each subagent’s functionality, input/output formats, and usage. This prompt-based onboarding allows any LLM-based agent to learn to use the ex- ported subagents without code-level modifications to the host framework. Problem Solving Once an external agent under- stands the available subagents through their code 1 https://github.com/anthropics/claude-code 4 query[LLM Parse JSON]Use path Hardcoded Fallback Run2: LLM JSON Parsing + Hardcoded Fallback (Fragile) success failure query[LLM Parse JSON]Use path Regex Fallback Run3: LLM JSON Parsing + Regex Fallback (Robust) success failure query[No Parsing]Use hardcoded path Run1: Hardcoded Path Figure 2: Evolution of path resolution mechanism across three runs. and SKILL.md descriptions, it can utilize multi- ple subagents to solve complex tasks—selecting and chaining relevant subagent scripts to address different parts of a problem directly. This positions AgentFactory not only as a self- evolving system but also as an agent factory that produces deployable agents for the AI ecosystem. 4 Self-Evolving Demonstration We demonstrate the self-evolving capability of AgentFactory through two concrete examples: (1) continuous subagent improvement through itera- tive refinement, and (2) subagent saving and cross- system reuse. 4.1 Iterative Refinement Figure 2 demonstrates self-evolution through a README generation subagent across three runs. In Run 1, the subagent was hardcoded for a specific project. In Run 2, it attempted dynamic LLM-based path parsing but used a fragile, hardcoded fallback on failure. In Run 3, the Meta-Agent replaced it with regex-based extraction, making parsing more robust. This shows the agent autonomously de- tecting limitations and improving error handling through iterative refinement. Each time a similar new task appeared, the Meta-Agent analyzed exe- cution feedback, identified weaknesses, and modi- fied the subagent code to make it more robust and general-purpose. 4.2Subagent Saving and Cross-System Reuse Figure 3 demonstrates the subagent saving and reuse mechanism across three trajectories in differ- ent agent systems. In Trajectory 1 (within Agent- Please help me complete the task in audio1.mp3 audio1.mp3: Please play ... using qqmusic. Subagent1: Audio Transcriber Subagent2: Music Player Please create a tencent document and write ... Please help me complete the task in audio2.mp3 Subagent3: Document Creator audio2.mp3: Please create a tencent document and write ... Step2: Create & Run Step3: Create & Run Step5: Create & Run Step1: Q1 Step4: Q2 Step7: Q3 Step6: Add Skills Step8: Run Subagent1 Success Step9: Run Subagent3 AgentFactory Other System Figure 3: Demonstration of subagent saving and direct reuse across three trajectories. Factory), when processing an audio task containing "Please play ... using Q Music", the Meta-Agent created two subagents: Audio Transcriber (for transcribing audio files to text) and Q Music Player (for playing music via Q Music). Both subagents were saved to the skill library. In Tra- jectory 2 (within AgentFactory), when asked to create a Tencent document, the system created and saved Document Creator subagent for creating and editing online spreadsheets. In Trajectory 3, the workflow migrated to Claude Code—a differ- ent agent system—which first learned how to use the saved subagents by reading the SKILL.md doc- umentation. When given a new problem (an audio file de- scribing a document creation task), the agent in Claude Code directly invoked the previously saved Audio Transcriber subagent to extract the task description. The transcription revealed: "Please create a Tencent document spreadsheet, fill the first row with ..." The agent then directly called the saved Document Creator subagent to execute this task, completing it successfully without needing to create new subagents. Together, these demonstrations illustrate the key values of AgentFactory: (1) the self-evolving capability enables continuous subagent improve- ment through feedback-driven iterative refinement, making subagents increasingly robust and general- purpose; (2) saved subagents can be directly reused to solve new problems, significantly improving problem-solving efficiency; and (3) saved sub- agents are portable across different agent systems, enabling cross-system capability transfer and de- ployment. 5 Evaluation To quantitatively evaluate the effectiveness of AgentFactory’s subagent saving and reuse mecha- 5 Avg. Tokens / Task MethodTask SettingOpus 4.6Sonnet 4.6 ReAct Batch 182986893 Batch 270227029 Self-Evolving Agents Batch 1 (from scratch)86088163 Batch 2 (w/ saved)62108223 AgentFactory Batch 1 (from scratch)43249199 Batch 2 (w/ saved)29713862 Table 1: Average output tokens per task across eval- uation configurations. Lower values indicate that the orchestrating agent requires less effort to complete tasks, reflecting more efficient subagent reuse. Token counts exclude subagent-internal LLM consumption. nism, we compare it against a self-evolving agent with the textual experience baseline and a ReAct baseline across multiple task settings and LLM backbones. 5.1 Experimental Setup Task Design. We curated two batches of tasks. Batch 1 consists of 15 real-world tasks spanning di- verse domains, including web information retrieval, data visualization, browser automation, and audio processing. All tasks require writing and executing Python code to produce outputs such as charts or analysis reports. Batch 2 consists of 15 additional tasks with similar structure but different specific requirements, serving as transfer evaluation targets. The complete list of tasks for both batches is pro- vided in Appendix B. Baselines.We consider two baselines. (1) ReAct: simply use ReAct(Yao et al., 2023) mode, where the agent solves each task from scratch without any accumulated knowledge. (2) Self-Evolving Agent (with Textual Experience): We adopt the same spirit from (Shinn et al., 2023; Zelikman et al., 2022), the agent writes and executes code for each task, and saves the resulting experience as textual summaries (e.g., what worked, what failed, and lessons learned). When solving subsequent tasks, the agent can query and retrieve relevant past expe- riences in textual form to inform its approach. Evaluation Metric.We report the average output token count of the orchestrating model per task, ex- cluding token consumption within subagent LLM calls. Since subagents function as tools in our framework, this metric isolates the orchestration- level effort and directly measures the efficiency gains from saving and reusing subagents. We en- sure that all tasks are completed without runtime errors in all experiments. Models. We evaluate with two LLM backbones: Claude Opus 4.6 2 and Claude Sonnet 4.6 3 . We consistently use the same LLM backbone for both the MetaAgent and subagents across all experi- ments. 5.2 Results Table 1 summarizes the results across all settings. In the table, "from scratch" represents that the evaluation starts without any saved subagents, and "w/ saved represents" running the second batch with subagents created when evaluating on the first batch. Subagent reuse significantly reduces orchestra- tion cost on Batch 2. When solving Batch 2 tasks with previously saved subagents, AgentFac- tory achieves substantially lower token consump- tion than both baselines. This confirms that reusing executable subagents is substantially more efficient than solving tasks from scratch or relying on textual experience summaries. Stronger models benefit from subagent saving even within Batch 1. A notable finding is that Opus 4.6 already shows a significant reduction in orchestration tokens during Batch 1 itself under AgentFactory (4324 vs. 8298 for ReAct), despite the 15 tasks in Batch 1 being from diverse do- mains with limited inter-task overlap. This sug- gests that even during the initial task-solving phase, a stronger model can recognize opportunities to reuse subagents created from earlier tasks in the same batch. This is a promising signal as founda- tion models continue to improve. 6 Conclusion We presented AgentFactory, a self-evolving frame- work with a three-phase pipeline. The Install phase constructs subagents from scratch to solve initial problems; the Self-Evolve phase detects limitations in saved subagents when encountering similar tasks and autonomously modifies them to be more robust and general-purpose; the Deploy phase exports ma- ture subagents as standalone Python code. This dual nature—both a self-evolving system and an agent factory—distinguishes AgentFactory from existing approaches. Through web-based tools, AgentFactory can interact with any application that 2 https://w.anthropic.com/news/claude-opus-4 -6 3 https://w.anthropic.com/claude/sonnet 6 has a web interface, enabling diverse task automa- tion. The unified skill interface maintains com- patibility with external frameworks. Future work will explore non-web applications by integrating Vision-Language Models (VLMs) to enable GUI- based interactions. 7 Ethical Considerations AgentFactory autonomously generates and exe- cutes code, which raises potential safety concerns. To mitigate risks, theshell_commandtool includes built-in security checks that flag destructive opera- tions before execution. Additionally, all saved sub- agents are stored as transparent, human-readable Python code with accompanying documentation, allowing users to inspect and audit any generated behavior before deployment. We also note that AgentFactory’s browser automation and web inter- action capabilities should only be used with proper authorization and in compliance with the terms of service of the accessed platforms. We encourage users to review exported subagents before integrat- ing them into production systems and to apply ap- propriate access controls when deploying agents that interact with external services. References Anthropic. 2026. Claude opus 4.6: Frontier intelligence for coding and agents.https://w.anthropic. com/news/claude-opus-4-6 . Accessed: 2026-03- 13. Harrison Chase. 2022. Langchain.https://github.c om/langchain-ai/langchain. Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2024. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors. In The Twelfth International Conference on Learning Representations. Chrisantha Fernando,Dylan Banarse,Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. 2024.Promptbreeder:Self-referential self- improvement via prompt evolution. In Forty-first International Conference on Machine Learning. Gemini team. 2026. Gemini 3.1 pro: A smarter model for your most complex tasks. Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yu- jiu Yang. 2024.EvoPrompt: Connecting llms with evolutionary algorithms yields powerful prompt opti- mizers. In The Twelfth International Conference on Learning Representations. Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024.MetaGPT: Meta programming for A multi-agent collaborative frame- work. In The Twelfth International Conference on Learning Representations. Shengran Hu, Cong Lu, and Jeff Clune. 2024. Au- tomated design of agentic systems. arXiv preprint arXiv:2408.08435. Zelong Li, Shuyuan Xu, Kai Mei, Wenyue Hua, Bal- aji Rama, Om Raheja, Hao Wang, He Zhu, and Yongfeng Zhang. 2024.AutoFlow: Automated work- flow generation for large language model agents. arXiv preprint arXiv:2407.12821. Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. 2023. DynamicLLM-Agentnetwork: An LLM-agent collaboration framework with agent team optimization. arXiv preprint arXiv:2310.02170. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems, 36:46534–46594. Alexander Novikov, Ngân V ̃ u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wag- ner, Sergey Shirobokov, Borislav Kozlovskii, Fran- cisco JR Ruiz, Abbas Mehrabian, and 1 others. 2025. AlphaEvolve: A coding agent for scientific and algo- rithmic discovery. arXiv preprint arXiv:2506.13131. OpenAI. 2026. Gpt-5.4 thinking system card. Technical report, OpenAI Deployment Safety Hub. Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2024.ChatDev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174–15186. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Re- flexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36:8634–8652. Significant Gravitas. 2023. Autogpt: Autonomous ai agents powered by large language models.https: //github.com/Significant-Gravitas/AutoGPT. 7 Xiangru Tang, Tianrui Qin, Tianhao Peng, Ziyang Zhou, Daniel Shao, Tingting Du, Xinming Wei, Peng Xia, Fang Wu, He Zhu, and 1 others. 2025. Agent kb: Leveraging cross-domain experience for agentic prob- lem solving. arXiv preprint arXiv:2507.06229. Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Man- dlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291. Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others. 2024a. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345. Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Hao- tian Luo, Jiayou Zhang, Nebojsa Jojic, Eric P. Xing, and Zhiting Hu. 2024b.PromptAgent: Strategic planning with language models enables expert-level prompt optimization. In The Twelfth International Conference on Learning Representations. Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, and 1 others. 2024. Au- togen: Enabling next-gen llm applications via multi- agent conversations. In First conference on language modeling. Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yi- wen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, and 1 others. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sci- ences, 68(2):121101. Wujiang Xu, Kai Mei, Hang Gao, Juntao Tan, Zu- jie Liang, and Yongfeng Zhang. 2025.A-MEM: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110. Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. 2023. Large language models as optimizers. In The Twelfth International Conference on Learning Representations. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. Preprint, arXiv:2210.03629. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Good- man. 2022. Star: Bootstrapping reasoning with rea- soning. Advances in Neural Information Processing Systems, 35:15476–15488. Barry Zhang, Keith Lazuka, and Mahesh Murag. 2025a. Equipping agents for the real world with agent skills. https://w.anthropic.com/engineering/eq uipping-agents-for-the-real-world-with-a gent-skills. Anthropic Engineering Blog, Oct 2025. Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. 2025b. Darwin godel machine: Open- ended evolution of self-improving agents. arXiv preprint arXiv:2505.22954. Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xiong-Hui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, and Chenglin Wu. 2025c. AFlow: Automating agentic workflow generation. In The Thirteenth International Conference on Learning Representations. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. 2025. MEM1: Learning to synergize memory and reason- ing for efficient long-horizon agents. arXiv preprint arXiv:2506.15841. Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. 2024. GPTSwarm: Language agents as optimizable graphs. In Forty-first International Conference on Machine Learning. 8 A Built-in Tools and Skills AgentFactory provides a comprehensive suite of built-in tools that can be used immediately or ex- tended. A.1 Meta Skills The core meta skills provide the fundamental opera- tions for agent orchestration and skill management: • create_subagent: Creates a new subagent with custom Python code and optional skill dependencies. The subagent is generated as executable code that can be immediately in- voked. • get_skill_description: Retrieves the doc- umentation of a specified skill, including its parameters and return format. • run_subagent : Executes a saved or newly created subagent with a given query. Handles the complete execution lifecycle including er- ror handling and result extraction. • modify_subagent: Modifies the code of an existing saved subagent based on new require- ments or feedback from execution. Enables iterative improvement of previously created skills. • finish: Completes the current task and op- tionally persists the developed subagents as reusable skills for future tasks. • list_saved_subagents: Lists all previously saved reusable skills in the skills directory, enabling the agent to discover and reuse accu- mulated capabilities. • view_subagent_code: Inspects the source code of any saved skill, allowing the agent to understand and modify existing implemen- tations. A.2 Tool Skills The built-in tool skills provide comprehensive ca- pabilities for diverse agent tasks: • web_search: Search capabilities via Serper API, enabling agents to retrieve relevant infor- mation from the web. • web_reading: Jina-based content extraction from web pages. • browser_automation:Playwright-based browser control for web interaction, support- ing persistent context for session management, dynamic element discovery, and LLM-driven automation loops. • shell_command : Safe command execution with security checks for destructive opera- tions. All skills follow a consistent interface with de- tailed descriptions of input/output formats, en- abling the Meta-Agent to intelligently select ap- propriate skills for each sub-task. B Evaluation Tasks Tables 2 and 3 list all 30 tasks used in the evaluation. Batch 1 is used for the first round evaluation, build- ing subagents from scratch, and Batch 2 serves as the transfer evaluation targets. Each Batch 2 task mirrors the structure of its Batch 1 counterpart but differs in specific requirements, enabling us to mea- sure subagent reuse efficiency. For the 13-th ques- tion, we use <audio_file_path> as a placeholder for the actual file path used during evaluation. 9 #Task Description 1Among various online discussions about the housing price bubble, how has public sentiment changed over time? Generate an analysis report and a trend chart, and include discussion and summary in the report. 2Search for and open the Stanford CS231n course homepage, find the syllabus or lecture list, and output the topics and corresponding links for the first 5 lectures. Save the final results to a markdown file. 3Book a Tencent Meeting for tomorrow at 4 PM with the topic “Work Meeting”. Find the meeting booking entry on the Tencent Meeting web version, then complete the booking through browser automation. Output the meeting invitation details, meeting ID, and meeting link. 4Search for China’s historical population data, then use matplotlib to plot the population change curve. 5Search for Bitcoin’s price data over the past 5 years, use matplotlib to plot the price trend, and calculate the annualized volatility. 6Search for and download trending keyword frequency data from a social media platform, and display the top 20 using a bar chart. 7 Search for and download China’s education expenditure and GDP data, plot a scatter chart showing the relationship between the two, and fit a regression line. 8Write a Python mini-game: Tetris, with keyboard controls and basic scoring logic. 9 Search for and open Wikipedia, find the “Transformer (machine learning model)” page, extract its core definition, and summarize the main ideas of Self-Attention. Save the results to a markdown file. 10Search for and open the GitHub website, find OpenAI’s official organization page, and output the page’s description and a list of main repositories (at least 5). Save the results to a markdown file. 11Search for and open the HuggingFace website, find a “Text-to-Image” model leaderboard or collection page, and summarize the 3 most common model architectures. Save the results to a markdown file. 12Create a soothing bedtime meditation audio file (15–20 min), featuring gentle guided breathing exercises, progres- sive muscle relaxation instructions, and calming visualizations. 13Complete a task whose detailed description is provided in <audio_file_path>. 14 Search travel resources about Tokyo. Under a budget of $100 USD, plan a 3-day itinerary with at least one cultural attraction per day, total travel time under 5 hours, and at least one anime-related location. Output the itinerary, budget breakdown, and total travel time. 15Search and download CO 2 emissions and renewable energy share data for at least 5 countries. Analyze the correlation, generate a plot, and summarize the findings. Table 2: Batch 1 evaluation tasks (15 tasks) used for initial subagent construction. #Task Description 1Among various online discussions about electric vehicle adoption, how has public sentiment changed over time? Generate an analysis report and a trend chart, and include discussion and summary in the report. 2 Search for and open the MIT 6.S191 (Introduction to Deep Learning) course homepage, find the syllabus or lecture list, and output the topics and corresponding links for the first 5 lectures. Save the final results to a markdown file. 3Book a Tencent Meeting for tomorrow at 7 PM with the topic “Work Meeting”. Find the meeting booking entry on the Tencent Meeting web version, then complete the booking through browser automation. Output the meeting invitation details, meeting ID, and meeting link. 4Search for Japan’s historical population data, then use matplotlib to plot the population change curve. 5 Search for Ethereum’s price data over the past 5 years, use matplotlib to plot the price trend, and calculate the annualized volatility. 6Search for and download trending topic data from a Chinese social media platform, and display the top 20 using a bar chart. 7Search for and download the US healthcare expenditure and GDP data, plot a scatter chart showing the relationship between the two, and fit a regression line. 8 Write a Python mini-game: Snake, with keyboard controls and basic scoring logic. 9Search for and open Wikipedia, find the “BERT (language model)” page, extract its core definition, and summarize the main ideas of the Transformer architecture. Save the results to a markdown file. 10Search for and open the GitHub website, find Meta AI’s official organization page, and output the page’s description and a list of main repositories (at least 5). Save the results to a markdown file. 11Search for and open the HuggingFace website, find a “Text-to-Speech” model leaderboard or collection page, and summarize the 3 most common model architectures. Save the results to a markdown file. 12Create a focus-enhancing background audio file (15–20 min), featuring gentle ambient sounds, subtle background music, and nature sounds to help concentrate. 13 Complete a task whose detailed description is provided in <audio_file_path>. 14Search travel resources about Paris. Under a budget of $150 USD, plan a 3-day itinerary with at least one museum per day, total travel time under 6 hours, and at least one historic landmark. Output the itinerary, budget breakdown, and total travel time. 15Search and download GDP per capita and life expectancy data for at least 5 countries. Analyze the correlation between economic development and health outcomes, generate a plot, and summarize the findings. Table 3: Batch 2 evaluation tasks (15 tasks) used as transfer evaluation targets. Each task mirrors the structure of its Batch 1 counterpart but differs in specific requirements. 10