# Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL

Source: [Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL](<https://www.sota2.com/research/paper/68fa7abc1e209131cea84e86>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-26T05:26:48.092Z

Authors: Weizhen Li, Jianbo Lin, Zhuosong Jiang, Jingyi Cao, Xinpeng Liu, Jiayu Zhang, Zhenqiang Huang, Qianben Chen, Weichen Sun, Qiexiang Wang, Hongxuan Lu, Tianrui Qin, Chenghao Zhu, Yi Yao, Shuying Fan, Xiaowan Li, Tiannan Wang, Pai Liu, King Zhu, He Zhu, Dingfeng Shi, Piaohong Wang, Yeyi Guan, Xiangru Tang, Minghao Liu, Yuchen Eleanor Jiang, Jian Yang, Jiaheng Liu, Ge Zhang, Wangchunshu Zhou

Published: 2025-08-06T17:01:02.000Z

## Abstract

Recent advances in large language models (LLMs) and multi-agent systems have demonstrated remarkable capabilities in complex problem-solving tasks such as deep research, vibe coding, and mathematical reasoning. However, most existing multi-agent systems are built upon manual prompt/workflow engineering with sophisticated agent frameworks, making them computationally inefficient, less capable, and can not benefit from data-centric learning. In this work, we introduce Chain-of-Agents (CoA), a novel paradigm of LLM reasoning that enables native end-to-end complex problem-solving in the same way as a multi-agent system (i.e., multi-turn problem solving with multiple tools and multiple agents) within one model. In chain-of-agents problem-solving, the model dynamically activates different tool agents and role-playing agents to simulate multi-agent collaboration in an end-to-end fashion. To elicit end-to-end chain-of-agents problem-solving abilities in LLMs, we introduce a multi-agent distillation framework to distill state-of-the-art multi-agent systems into chain-of-agents trajectories for agentic supervised fine-tuning. We then use agentic reinforcement learning on verifiable agentic tasks to further improve the models' capabilities on chain-of-agents problem solving. We call the resulting models Agent Foundation Models (AFMs). Our empirical studies demonstrate that AFM establishes new state-of-the-art performance across diverse benchmarks in both web agent and code agent settings. We make the entire research, including the model weights, code for training and evaluation, and the training data, fully open-sourced, which offers a solid starting point for future research on agent models and agentic RL.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Code Generation on CodeContests (Accuracy)](<https://www.sota2.com/research/sota/code-generation-on-codecontests-accuracy.md>): Code Generation; CodeContests; Accuracy: 21.2.

- [Code Generation on LiveCodeBench V4 (test)](<https://www.sota2.com/research/sota/code-generation-on-livecodebench-v4-test.md>): Code Generation; LiveCodeBench V4 (test); pass@1: 44.9.

- [Multi-hop reasoning and question answering on GAIA](<https://www.sota2.com/research/sota/multi-hop-reasoning-and-question-answering-on-gaia.md>): Multi-hop reasoning and question answering; GAIA; GAIA L1 Acc: 69.2.

- [Multi-hop reasoning and question answering on HLE](<https://www.sota2.com/research/sota/multi-hop-reasoning-and-question-answering-on-hle.md>): Multi-hop reasoning and question answering; HLE; HLE Average: 18.

- [Multi-hop Reasoning and Question Answering on PopQA](<https://www.sota2.com/research/sota/multi-hop-reasoning-and-question-answering-on-popqa.md>): Multi-hop reasoning and question answering; PopQA; PopQA Score: 46.5.

- [Web Navigation Question Answering on WebWalker QA (avg@3)](<https://www.sota2.com/research/sota/web-navigation-question-answering-on-webwalker-qa-avg-at-3.md>): Web Navigation Question Answering; WebWalker QA; Avg@3: 63.

- [Code Generation on HumanEval (test)](<https://www.sota2.com/research/sota/code-generation-on-humaneval-test.md>): Code Generation; HumanEval (test); Pass@1: 95.3.

- [Code Generation on MBPP (test)](<https://www.sota2.com/research/sota/code-generation-on-mbpp-test.md>): Code Generation; MBPP (test); Pass@1: 90.2.

- [Code Generation on APPS (Introductory, Interview, Competition)](<https://www.sota2.com/research/sota/code-generation-on-apps-introductory-interview-competition.md>): Code Generation; APPS; Accuracy: 28.

- [Code Generation on CodeContests (test)](<https://www.sota2.com/research/sota/code-generation-on-codecontests-test.md>): Code Generation; CodeContests (test); Pass@1: 34.6.

- [Deep-search QA on BrowseComp (test)](<https://www.sota2.com/research/sota/deep-search-qa-on-browsecomp-test.md>): Deep-search QA; BrowseComp (test); Pass@1: 11.1.

- [Reasoning Question Answering on FinQA (test)](<https://www.sota2.com/research/sota/reasoning-question-answering-on-finqa-test.md>): Reasoning Question Answering; FinQA (test); Precision@1: 65.1.

- [Web Browsing and Navigation on BrowseComp](<https://www.sota2.com/research/sota/web-browsing-and-navigation-on-browsecomp.md>): Web Browsing and Navigation; Browsecomp; Avg@3 Score: 11.1.

- [Autonomous Agent Problem Solving on Humanity's Last Exam](<https://www.sota2.com/research/sota/autonomous-agent-problem-solving-on-humanity-s-last-exam.md>): Autonomous Agent Problem Solving; Humanity's Last Exam; Avg@3: 18.

- [Web Task Reasoning on XBench (test)](<https://www.sota2.com/research/sota/web-task-reasoning-on-xbench-test.md>): Web Task Reasoning; xbench (test); Pass@1: 21.2.

- [Tabular Reasoning on MultiHier (test)](<https://www.sota2.com/research/sota/tabular-reasoning-on-multihier-test.md>): Tabular Reasoning; MultiHier (test); Pass@1: 44.9.

- [Web Task Reasoning on GAIA (test)](<https://www.sota2.com/research/sota/web-task-reasoning-on-gaia-test.md>): Web Task Reasoning; GAIA (test); Pass@1: 22.6.

- [Scientific Reasoning on HLE (test)](<https://www.sota2.com/research/sota/scientific-reasoning-on-hle-test.md>): Scientific Reasoning; HLE (test); Pass@1: 7.8.

- [Web Task Reasoning on BrowseComp (test)](<https://www.sota2.com/research/sota/web-task-reasoning-on-browsecomp-test.md>): Web Task Reasoning; BrowseComp (test); Pass@1: 1.2.

- [General AI Assistant Tasks on GAIA (avg@8)](<https://www.sota2.com/research/sota/general-ai-assistant-tasks-on-gaia-avg-at-8.md>): General AI Assistant Tasks; GAIA; Avg@8 Score: 55.3.

- [Scientific Reasoning on SuperGPQA (test)](<https://www.sota2.com/research/sota/scientific-reasoning-on-supergpqa-test.md>): Scientific Reasoning; SuperGPQA (test); Pass@1: 32.4.

- [Mathematical Reasoning on AIME25 (test)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-aime25-test.md>): Mathematical Reasoning; AIME25 (test); Pass@1: 30.1.

- [Mathematical Reasoning on AIME24 (test)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-aime24-test.md>): Mathematical Reasoning; AIME24 (test); Pass@1 Score: 39.7.

- [Code Generation on LiveCodeBench (Accuracy and Average Score)](<https://www.sota2.com/research/sota/code-generation-on-livecodebench-accuracy-and-average-score.md>): Code Generation; LiveCodeBench; Accuracy: 29.1.

- [General Reasoning on General Reasoning Suite Average](<https://www.sota2.com/research/sota/general-reasoning-on-general-reasoning-suite-average.md>): General Reasoning; General Reasoning Suite Average; Pass@1: 28.9.

- [Mathematical Reasoning on AIME 25 (Acc@8, Pass@8)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-aime-25-acc-at-8-pass-at-8.md>): Mathematical Reasoning; AIME 25; Accuracy: 8.

- [Mathematical Reasoning on AIME 24 (Accuracy, Tokens)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-aime-24-accuracy-tokens.md>): Mathematical Reasoning; AIME 24; Accuracy: 12.

- [Mathematical Reasoning on MATH500 (test)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-math500-test.md>): Mathematical Reasoning; MATH500 (test); Accuracy: not recorded.

- [Code Generation on APPS (test)](<https://www.sota2.com/research/sota/code-generation-on-apps-test.md>): Code Generation; APPS (test); Introductory Score: not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2508.13167>)

[PDF](<https://arxiv.org/pdf/2508.13167.pdf>)
