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

Zhang Zhang; Shuqi Lu; Hongjin Qian; Di He; Zheng Liu

cs.AIMar 18, 2026

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

Authors: Zhang Zhang, Shuqi Lu, Hongjin Qian, Di He, Zheng Liu

Organizations: Peking University · Beijing Academy of Artificial Intelligence · Hong Kong Polytechnic University

Abstract

Building LLM-based agents has become increasingly important. Recent works on LLM-based agent self-evolution primarily record successful experiences as textual prompts or reflections, which cannot reliably guarantee efficient task re-execution in complex scenarios. We propose AgentFactory, a new self-evolution paradigm that preserves successful task solutions as executable subagent code rather than textual experience. Crucially, these subagents are continuously refined based on execution feedback, becoming increasingly robust and efficient as more tasks are encountered. Saved subagents are pure Python code with standardized documentation, enabling portability across any Python-capable system. We demonstrate that AgentFactory enables continuous capability accumulation: its library of executable subagents grows and improves over time, progressively reducing the effort required for similar tasks without manual intervention. Our implementation is open-sourced at https://github.com/zzatpku/AgentFactory, and our demonstration video is available at https://youtu.be/iKSsuAXJHW0.

Figures & tables

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.

Figure 2: Evolution of path resolution mechanism across three runs.

Figure 3: Demonstration of subagent saving and direct reuse across three trajectories.

		Avg. Tokens / Task
Method	Task Setting	Opus 4.6	Sonnet 4.6
ReAct	Batch 1	8298	6893
ReAct	Batch 2	7022	7029
Self-Evolving Agents	Batch 1 (from scratch)	8608	8163
Self-Evolving Agents	Batch 2 (w/ saved)	6210	8223
AgentFactory	Batch 1 (from scratch)	4324	9199

Table 1: Average output tokens per task across evaluation 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.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

#	Task Description
1	Among 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.
2	Search 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.
3	Book 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.
4	Search for China’s historical population data, then use matplotlib to plot the population change curve.
5	Search for Bitcoin’s price data over the past 5 years, use matplotlib to plot the price trend, and calculate the annualized volatility.
6	Search for and download trending keyword frequency data from a social media platform, and display the top 20 using a bar chart.

Appendix

Table 2: Batch 1 evaluation tasks (15 tasks) used for initial subagent construction.

#	Task Description
1	Among 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.
3	Book 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.
4	Search 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.
6	Search for and download trending topic data from a Chinese social media platform, and display the top 20 using a bar chart.

Appendix

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.

Self-Evolving Agents

Self-evolving agents improve their skills, memory, tools, policies, or environments by learning from interaction and evaluation feedback.

See all 421 papers

Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses
Aug 9, 2026Tailin ZhouAgent Harness Optimization LLM Agent Self-Improvement
HKUST
Mem $^2$ Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation
Apr 13, 2026Zihao Cheng, Zeming Liu, Yingyu Shan +7LLM Agent Self-Improvement Self-Evolving Agents
Beihang University · Beijing Institute of Technology · Independent Researcher +2
EXG: Self-Evolving Agents with Experience Graphs
May 18, 2026Yuxin Jin, Siyuan Zhang, Hanchen Wang +3Graph-Based Agent Memory Memory-Augmented Agents
University of Technology Sydney Sydney, Australia · The University of New South Wales Sydney, Australia

Aug 9, 2026J/K move · Enter open · S save

Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

Tailin Zhou

HKUST

Agent Harness Optimization LLM Agent Self-Improvement LLM Agents

Open paper AskPDF arXiv

Modern LLM agents are often improved by modifying prompts, tools, or workflows manually, while the executable scaffold surrounding the model---the harness---is typically treated as a fixed artifact after deployment. This work studies an alternative where the harness is task-specific and continuously evolvable: each task family maintains its own harness, which is hot-swapped across iterations through a fixed task-injection seam and rewritten using environment feedback. We introduce Hierarchical Self-Improvement (HSI), a framework in which a single frozen LLM

M

operates across three hierarchical scopes: a task harness

H

that executes tasks, an evolver that rewrites

H

, and a meta-evolver that rewrites the evolver's strategy code under a frozen outer anchor. A thinking-on/off design isolates the contribution of harness evolution by disabling reasoning during task execution while enabling it during self-modification. HSI is bounded by two factors: a feedback-fidelity bound, since evolution requires informative reward signals to guide selection, and a backbone capability bound, since harness redesign cannot overcome limitations of the frozen model. On BALROG with DeepSeek-V4-Flash-Preview as the frozen backbone, HSI achieves consistent gains over the initial harness on moderate-difficulty tasks (

+39.3

on BabyAI,

+33.0

on Crafter,

+25.0

on TextWorld, and

+15.0

on MiniHack, all in raw % Progress), while obtaining strong held-out generalization on BabaIsAI sub-suites (

0.98

best-test on BreakStop and

1.00

on GoTo from a

20\%

unseen split). On tasks beyond the backbone's capability (NLE), harness evolution provides no improvement. These results demonstrate task-specific harness evolution as a viable axis for improving frozen LLM agents under clear empirical limits. Code is available at https://github.com/TailinZhou/hsi.

Abstract

Self-Evolving Agents

Explore similar work

Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

Mem2^22Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation

EXG: Self-Evolving Agents with Experience Graphs

Mem $^2$ Evolve: Towards Self-Evolving Agents via Co-Evolutionary Capability Expansion and Experience Distillation