Aime: Towards Fully-Autonomous Multi-Agent Framework

Chen Li; Hongjie Lai; Jian Xu; Jie Geng; Junjian Lan; Mingyu Wang; Wenbo Duan; Xiang Chen; Xin Han; Yexuan Shi

REVIEW 5 major objections 5 minor 3 cited by

Aime: Towards Fully-Autonomous Multi-Agent Framework

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims one adaptive multi-agent framework outperforms specialized agents on general reasoning, software engineering, and live web navigation.

desk verdict A clear and sensible framework architecture, but the SOTA claims are not supported by the experimental controls as written; worth sending to review, not worth citing yet. read the letter →

arxiv 2507.11988 v2 pith:XDVM46MG submitted 2025-07-16 cs.AI

Yexuan Shi , Mingyu Wang , Yunxiang Cao , Hongjie Lai , Junjian Lan , Xin Han , Yu Wang , Jie Geng

show 7 more authors

Zhenan Li Zihao Xia Xiang Chen Chen Li Jian Xu Wenbo Duan Yuanshuo Zhu

This is my paper

classification cs.AI

keywords multi-agentsystemsLLMagentsdynamicplanningactorfactoryprogressmanagementGAIAbenchmarkSWE-benchVerifiedWebVoyager

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that the dominant plan-and-execute design for multi-agent LLM systems is the main obstacle to robustness, and that replacing it with a continuously reactive loop makes one general framework beat specialists in three different domains. It introduces Aime, whose Dynamic Planner revises the task plan from live feedback, whose Actor Factory builds a purpose-made agent for each subtask, and whose Progress Management Module keeps a shared, current task list as the single source of truth. Evaluated on GAIA, SWE-bench Verified, and WebVoyager, Aime reports success rates of 77.6%, 66.4%, and 92.3%, each above the specialized baselines the paper compares against. If these numbers hold, a single architecture could replace domain-specific agent designs.

What carries the argument

The central object is the progress list: a hierarchical, human-readable task tree (a Markdown checklist) that encodes subtasks, completion criteria, and dependencies, and acts as the single source of truth for the whole system. Around it sit three components: the Dynamic Planner, which at each step takes the goal, the current list, and the history of outcomes and outputs both an updated global list and the next concrete action; the Actor Factory, which on demand assembles an actor from a persona, a prepackaged tool bundle, retrieved knowledge, environment context, and an output schema; and the Dynamic Actor, which executes subtasks via the ReAct reasoning-action-observation loop and can push progress updates through a special Update Progress tool. The design's claim is that this loop removes the three failure modes it attributes to plan-and-execute: rigid plans, static roles, and context loss at handoff.

What would settle it

Re-running Aime and the baselines on the same fresh splits of GAIA, SWE-bench Verified, and WebVoyager with one named LLM and an identical harness; if Aime's margins disappear or reverse under controlled conditions, the framework-level advantage is not established.

Watch

Extended reading notes

Core claim

In the paper's own framing, the discovery is that the bottleneck in contemporary multi-agent LLM systems is architectural, not a matter of model scale or tool count. Aime replaces the one-shot plan-and-execute pipeline with a closed loop in which the planner continuously rewrites a global task list, the Actor Factory instantiates a fresh specialized agent (persona, tool bundle, knowledge, output format) for each subtask, and a centralized progress list gives every component the same real-time state. Each agent runs the ReAct cycle and can autonomously call an Update Progress tool, so the planner reacts mid-task instead of after completion. The paper reports that this design yields 77.6% on GAIA, 66.4% on SWE-bench Verified, and 92.3% on WebVoyager, establishing, in its account, a new state of the art in all three domains with a single framework.

Load-bearing premise

The paper assumes its reported scores and the baselines' scores were obtained under comparable conditions, specifically the same underlying LLM and the same evaluation protocol, so that the gap is attributable to the framework.

Editorial extensions

If this is right

A single Aime-style framework could replace domain-specific agents for mixed workloads, since the paper reports state-of-the-art results on general reasoning, code repair, and live web tasks with the same architecture.
The progress list makes task status auditable at any instant, so a system built on it can be paused, resumed, or redirected mid-task without losing context.
New capabilities can be added by registering a new tool bundle or knowledge source, rather than re-engineering a fixed roster of agent roles.
The reported GAIA gain is attributed to the Dynamic Planner's re-planning on failure, implying that reactive re-planning is a general lever for hard multi-step problems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: because the paper does not name the underlying LLM, the headline gaps could partly reflect model strength rather than architecture; a same-model, same-harness replication would settle whether the framework itself carries the gain.
Editorial inference: the Actor Factory's bundle-based tool selection implies a governance question—who curates the bundles and how errors propagate when a bundle omits a needed tool—that the paper does not address.
Editorial inference: the progress list format suggests a natural stress test: measure token cost and wall-clock time against the number of subtasks, since a centralized list may become a coordination bottleneck for very large agent teams.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a plausible multi-agent framework with a clean, well-motivated design, but the empirical section does not back the headline claims. The stress-test concern is right: the base LLM is never named, the baselines are mostly numbers lifted from other papers or closed products, and there are no re-runs, error bars, or run counts. The SWE-bench margin is 0.6 points with no variance, which is noise-level. The GAIA split is also ambiguous — 'public test set' could easily mean the validation set.

What is genuinely new: the combination of a Dynamic Planner that re-plans from live feedback, an Actor Factory that instantiates specialized actors on demand with tool bundles and generated personas, and a centralized progress list as a single source of truth. The pieces are familiar (ReAct, hierarchical task lists, tool bundles), but the integration is a reasonable step beyond static plan-and-execute, and the workflow is described clearly enough that a reader can understand what the system does.

The soft spots are exactly the ones the stress-test names, in proportion: the experiments are the entire basis for the 'state of the art' claim, and they are not reproducible from the paper. No LLM, no sampling details, no error bars. The baselines are not re-run in the same harness, so differences could be due to model strength or evaluation protocol rather than the framework. The closed-product baselines (Operator, Skyvern) cannot be matched at all. There are also no ablations isolating the Dynamic Planner or Actor Factory; the 77.6% GAIA result is attributed to the planner without evidence. The circularity burden is low — no parameters are fitted to the benchmarks — but the component-level attribution is unsupported.

None of this invalidates the architecture as a design proposal. A proper revision needs: a named base LLM, re-runs of at least the open baselines under the same evaluation code, error bars or run counts, a clear statement of which GAIA split was used, and ablations that turn off dynamic planning and the actor factory. Code or a detailed appendix would help enormously.

For peer review: I would send it out. The idea is significant enough, and a referee can productively push on the experimental claims. But I would not cite the numbers as state of the art until they are independently reproduced. The paper is worth a reading group discussion mostly about what counts as a fair comparison in this rapidly moving area.

Referee Report

5 major / 5 minor

Summary. The paper introduces Aime, a multi-agent framework that replaces static plan-and-execute workflows with a dynamic planner, an actor factory that instantiates specialized actors on demand, and a centralized progress management module. The authors report state-of-the-art results on three benchmarks: 77.6% on GAIA, 66.4% on SWE-bench Verified, and 92.3% on WebVoyager, claiming that a single general framework outperforms specialized agents in each domain. The architectural description is detailed and includes formal equations for the planner, the actor factory, and the actor execution loop, followed by a brief experimental section and related work.

Significance. The architectural ideas are timely and plausible: reactive planning, dynamic role creation, and centralized state management directly address recognized weaknesses of fixed multi-agent teams. If the empirical claims were supported by controlled comparisons, the paper would make a strong contribution by demonstrating that one general framework can match or beat specialized agents on multiple benchmarks. However, the evaluation as presented does not provide controlled evidence. The underlying LLM is never named, no error bars or repeated runs are reported, and baseline numbers appear to be taken from prior papers or vendor web pages rather than produced in a shared harness. Consequently, the central state-of-the-art claim is not verifiable and may reflect model choice, evaluation protocol, or task subset rather than the value of the proposed framework.

major comments (5)

[§5.1, Baselines] Section 5.1 states that "all agents, including our own, are powered by the same underlying LLM, where applicable," but the LLM is never named anywhere in the paper. This is a load-bearing confound: the reported performance gaps between Aime and the baselines could be entirely due to a stronger base model rather than to the Dynamic Planner, Actor Factory, or Progress Management Module. The authors must specify the exact LLM, version, and configuration used for Aime and for each baseline, or explicitly disclose which baselines did not use the same model.
[§5.1 and Table 1] The baseline results in Table 1 appear to be cited from external sources rather than generated in a common evaluation harness. For GAIA, the baselines (Langfun 71.5, Trase 70.3, OWL 69.1) and for SWE-bench (SWE-agent 62.4, OpenHands 65.8) are single point estimates with no error bars. On WebVoyager, Operator and Skyvern are closed products whose underlying models cannot be controlled by the authors. Without re-running the open baselines under the same model and evaluation protocol, and without acknowledging the limitations of closed-product comparisons, the claim that Aime "consistently outperforms" specialized state-of-the-art agents is not supported.
[§5.1, GAIA evaluation] The paper says Aime is evaluated on the "public test set" of GAIA, but GAIA's official test set is private; the public split is the validation set. If the authors used the validation set instead, the comparison to baselines that may have been evaluated on the private test set is invalid. Please clarify exactly which split was used and confirm that all compared methods are measured on the identical split and with the same evaluation script and exact-match metric.
[Table 1 and §5.2] No measure of variance, number of runs, or statistical significance is reported. The difference on SWE-bench Verified is 66.4% versus OpenHands's 65.8%, a gap of 0.6 percentage points that is likely within run-to-run noise for a benchmark of this type. Repeated runs with confidence intervals, or at least a clear statement that the difference is within noise, are necessary before the claim "Aime resolves more issues than top specialized agents" can be accepted.
[§4 and §5.2] The paper attributes performance to specific components without any ablation study. For example, the WebVoyager result is credited to the "tight feedback loop" between Dynamic Actors and the Dynamic Planner, and the SWE-bench result is credited to the Actor Factory instantiating different roles on the fly, but no experiment isolates these components. Without ablations or controlled variations, the causal role of the proposed mechanisms in the reported performance is unsubstantiated.

minor comments (5)

[§5.1] The phrase "where applicable" in the statement about the same underlying LLM is ambiguous. It should either be removed if all agents truly use the same model or expanded to explain exactly which baselines are excluded and why.
[§4.1–4.3] The formal definitions in Equations (1)–(4) would benefit from more precise notation; for example, Eq. (2) does not describe how the Actor Factory selects from the available tool bundles or knowledge modules, leaving the core contribution underspecified.
[§8, References] Some references are cited only by arXiv or URL without page numbers or venue details, and the reference list style is inconsistent. Please ensure all entries follow a uniform format.
[§6.1] The related work cites MarsCode Agent (Liu et al., 2024), which includes one of the current authors. The citation is likely legitimate, but the self-citation should be contextualized as prior work distinct from Aime rather than presented as an external system.
[§5] The experimental section does not mention the number of evaluation runs, the compute configuration, or the temperature/sampling settings. Reporting these details is essential for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: Aime's benchmark claims are empirical comparisons, not outputs derived from fitted inputs or self-cited constraints.

full rationale

Aime does not present a derivation chain in which a result is constructed from its own inputs. The framework equations (1)-(4) formalize planner and actor state transitions, but no parameter is fitted to GAIA, SWE-bench Verified, or WebVoyager and then reported as a prediction; the success rates are measured on external benchmarks against cited baselines. The only apparent self-citation is MarsCode Agent (Liu et al., 2024) in the Related Work section, which includes one co-author overlap (Yexuan Shi); it is used only as an example of static role-based MAS and does not justify Aime's design choices or benchmark claims, so it is not load-bearing. The sentence 'all agents, including our own, are powered by the same underlying LLM, where applicable' is a comparability caveat; if baseline conditions were not actually controlled, that is an empirical correctness risk, not circular reasoning. No uniqueness theorem, ansatz-smuggled-via-citation, or renaming of a known result was found. The paper is self-contained against external benchmarks, so the circularity score is kept at the minimal non-zero level to acknowledge the minor self-citation without treating it as load-bearing.

Assumptions & free parameters 0 free parameters · 3 assumptions · 3 invented entities

The framework introduces three new components (Actor Factory, Progress Management Module, Dynamic Actor) but provides no external evidence for their individual contributions. The central claims rest on domain assumptions about the LLM, the benchmarks, and the comparability of baselines.

assumptions (3)

domain assumption The underlying LLM is capable enough to follow the prompt templates and tool-use protocols.
The entire framework relies on the LLM's instruction following and tool-use ability, which is assumed but not tested.
domain assumption The benchmarks and their official metrics measure the claimed capabilities.
The paper relies on GAIA, SWE-bench Verified, and WebVoyager as valid measures of general reasoning, software engineering, and web navigation, respectively.
domain assumption The cited baseline numbers are accurate and were obtained under comparable conditions.
The paper compares to numbers from prior papers without re-running baselines, assuming they are directly comparable.

invented entities (3)

Actor Factory
purpose: On-demand instantiation of specialized agents with selected personas, tools, and knowledge.
No external validation or falsifiable handle is provided; its effectiveness is only asserted through the paper's own results.
Progress Management Module
purpose: Centralized state tracking and coordination across the planner and actors.
The claimed benefit of reduced information loss is not independently confirmed; no ablation separates its effect.
Dynamic Actor
purpose: ReAct-based executor with autonomous progress reporting.
The architecture is described, but no code or external test isolates its contribution.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aime: Towards Fully-Autonomous Multi-Agent Framework." pith.science (2026). https://pith.science/paper/XDVM46MG

@misc{pith2026250711988,
  author       = {Pith},
  title        = {Pith review of: Aime: Towards Fully-Autonomous Multi-Agent Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XDVM46MG}},
  note         = {Machine review of arXiv:2507.11988}
}

read the original abstract

Multi-Agent Systems (MAS) powered by Large Language Models (LLMs) are emerging as a powerful paradigm for solving complex, multifaceted problems. However, the potential of these systems is often constrained by the prevalent plan-and-execute framework, which suffers from critical limitations: rigid plan execution, static agent capabilities, and inefficient communication. These weaknesses hinder their adaptability and robustness in dynamic environments. This paper introduces Aime, a novel multi-agent framework designed to overcome these challenges through dynamic, reactive planning and execution. Aime replaces the conventional static workflow with a fluid and adaptive architecture. Its core innovations include: (1) a Dynamic Planner that continuously refines the overall strategy based on real-time execution feedback; (2) an Actor Factory that implements Dynamic Actor instantiation, assembling specialized agents on-demand with tailored tools and knowledge; and (3) a centralized Progress Management Module that serves as a single source of truth for coherent, system-wide state awareness. We empirically evaluated Aime on a diverse suite of benchmarks spanning general reasoning (GAIA), software engineering (SWE-bench Verified), and live web navigation (WebVoyager). The results demonstrate that Aime consistently outperforms even highly specialized state-of-the-art agents in their respective domains. Its superior adaptability and task success rate establish Aime as a more resilient and effective foundation for multi-agent collaboration.

Figures

Figures reproduced from arXiv: 2507.11988 by the authors.

**Figure 1.** The workflow of Aime framework. Dynamic Planner. The Dynamic Planner serves as the central orchestrator for task management. It decomposes high-level objectives into a hierarchical structure of executable subtasks and maintains a global task list that tracks the status of each subtask. It continuously monitors execution progress and dynamically adapts the plan based on feedback from Dynamic Actors and status updat… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

MemEvolve: Meta-Evolution of Agent Memory Systems
cs.CL 2025-12 unverdicted novelty 7.0 of 10

MemEvolve jointly evolves agent experiential knowledge and memory architectures via a modular codebase, delivering up to 17% gains on agent benchmarks with cross-task and cross-model generalization.
Complete Cyclic Subtask Graphs for Tool-Using LLM Agents: Flexibility, Cost, and Bottlenecks in Multi-Agent Workflows
cs.MA 2026-04 unverdicted novelty 6.0 of 10

Complete cyclic subtask graphs offer a lens to measure when multi-agent revisitation aids recovery and exploration versus when it increases costs or is dominated by other bottlenecks in LLM agent workflows.
OxyGent: Making Multi-Agent Systems Modular, Observable, and Evolvable via Oxy Abstraction
cs.AI 2026-04 unverdicted novelty 4.0 of 10

OxyGent supplies a modular framework for multi-agent systems via the Oxy abstraction for composition and monitoring and the OxyBank engine for continuous automated evolution.

Reference graph

Works this paper leans on

13 extracted references · 2 canonical work pages · cited by 3 Pith papers

[1]

Language models are few-shot learners

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott G...

work page 1901
[4]

Agentverse: Fa- cilitating multi-agent collaboration and exploring emergent behaviors in agents

Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. Agentverse: Fa- cilitating multi-agent collaboration and exploring emergent behaviors in agents. ArXiv Preprint, arXiv:2308.10848,

arXiv
[6]

OWL: optimized workforce learning for general multi-agent assistance in real-world task automation

Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, Zeyu Zhang, Yifeng Wang, Qianshuo Ye, Bernard Ghanem, Ping Luo, and Guohao Li. OWL: optimized workforce learning for general multi-agent assistance in real-world task automation. ArXiv Preprint, arXiv:2505.23885,

arXiv
[8]

Understanding the planning of LLM agents: A survey

Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. Understanding the planning of LLM agents: A survey. ArXiv Preprint, arXiv:2402.02716,

arXiv
[9]

Marscode agent: Ai-native automated bug fixing

Yizhou Liu, Pengfei Gao, Xinchen Wang, Jie Liu, Yexuan Shi, Zhao Zhang, and Chao Peng. Marscode agent: Ai-native automated bug fixing. ArXiv Preprint, arXiv:2409.00899,

arXiv
[10]

GPT-4 technical report

OpenAI. GPT-4 technical report. ArXiv Preprint, arXiv:2303.08774,

arXiv
[11]

V oyager: An open-ended embodied agent with large language models

Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024, 2024a. Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne...

work page 2024
[12]

Autogen: Enabling next-gen LLM applications via multi-agent conversation framework

13 Aime Team Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen LLM applications via multi-agent conversation framework. ArXiv Preprint, arXiv:2308.08155,

arXiv

Show all 13 references

[13]

Multi-agent architecture search via agentic supernet

Guibin Zhang, Luyang Niu, Junfeng Fang, Kun Wang, Lei Bai, and Xiang Wang. Multi-agent architecture search via agentic supernet. ArXiv Preprint, arXiv:2502.04180, 2025a. Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, ...

arXiv
[2020]

Pan, Shuyi Yang, Lakshya A

Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Ti- wari, Kurt Keutzer, Aditya G. Parameswaran, Dan Klein, Kannan Ramchandran, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica. Why do multi-agent LLM systems fail? ArXiv Preprint, arXiv:2503.13657,

arXiv
[2023]

Flowreasoner: Reinforcing query-level meta-agents

Hongcheng Gao, Yue Liu, Yufei He, Longxu Dou, Chao Du, Zhijie Deng, Bryan Hooi, Min Lin, and Tianyu Pang. Flowreasoner: Reinforcing query-level meta-agents. ArXiv Preprint, arXiv:2504.15257,

arXiv
[2024]

Coder: Issue resolving with multi-agent and task graphs

Dong Chen, Shaoxin Lin, Muhan Zeng, Daoguang Zan, Jian-Gang Wang, Anton Cheshkov, Jun Sun, Hao Yu, Guoliang Dong, Artem Aliev, Jie Wang, Xiao Cheng, Guangtai Liang, Yuchi Ma, Pan Bian, Tao Xie, and Qianxiang Wang. Coder: Issue resolving with multi-agent and task graphs. ArXiv ...

arXiv
[2025]

Automated design of agentic systems

Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. ArXiv Preprint, arXiv:2408.08435,

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Language models are few-shot learners

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott G...

work page 1901

[2] [4]

Agentverse: Fa- cilitating multi-agent collaboration and exploring emergent behaviors in agents

Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. Agentverse: Fa- cilitating multi-agent collaboration and exploring emergent behaviors in agents. ArXiv Preprint, arXiv:2308.10848,

arXiv

[3] [6]

OWL: optimized workforce learning for general multi-agent assistance in real-world task automation

Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, Zeyu Zhang, Yifeng Wang, Qianshuo Ye, Bernard Ghanem, Ping Luo, and Guohao Li. OWL: optimized workforce learning for general multi-agent assistance in real-world task automation. ArXiv Preprint, arXiv:2505.23885,

arXiv

[4] [8]

Understanding the planning of LLM agents: A survey

Xu Huang, Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. Understanding the planning of LLM agents: A survey. ArXiv Preprint, arXiv:2402.02716,

arXiv

[5] [9]

Marscode agent: Ai-native automated bug fixing

Yizhou Liu, Pengfei Gao, Xinchen Wang, Jie Liu, Yexuan Shi, Zhao Zhang, and Chao Peng. Marscode agent: Ai-native automated bug fixing. ArXiv Preprint, arXiv:2409.00899,

arXiv

[6] [10]

GPT-4 technical report

OpenAI. GPT-4 technical report. ArXiv Preprint, arXiv:2303.08774,

arXiv

[7] [11]

V oyager: An open-ended embodied agent with large language models

Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024, 2024a. Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne...

work page 2024

[8] [12]

Autogen: Enabling next-gen LLM applications via multi-agent conversation framework

13 Aime Team Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen LLM applications via multi-agent conversation framework. ArXiv Preprint, arXiv:2308.08155,

arXiv