WeChat (微信)

English | 中文

The First Unified Agent Data Synthesis Framework for Custom Task with all-in-one envrionment.

🚀 Overview

AgentFlow is the first unified agent data synthesis framework that generates high-quality training and evaluation data across heterogeneous agent environments — including

• 📚 RAG

• 🖼️ MM-Doc

• 🔍 Deep Research

• 🖱️ GUI

• 🟰 Text2SQL

• 📊 Data Analysis

• 🤖 Embodied Agents

• and more.

It provides a unified, extensible, all-in-one environment for synthesizing agent trajectories, reasoning traces, tool interactions, and environment feedback.

AgentFlow also explores the underlying mechanisms of agent data synthesis and model training, enabling the development of industrial-grade agentic foundation models that operate seamlessly across domains.

Beyond synthetic training data, AgentFlow also offers high-quality human-annotated and synthetic benchmarks for evaluating emerging agent capabilities and exploring their boundaries.

One framework. All agent worlds.

✨ Key Features

Unified Agent Data Synthesis Paradigm

• Synthesize complex agent training data with just a few lines of code.
• Provide a unified abstraction layer for seamless data synthesis across heterogeneous agent environments.

All-in-One Sandbox

• Built-in support for 📚 RAG, 🖼️ MM-Doc, 🔍 Deep Research, 💻 Code, 🟰 SQL database, 🖱️ GUI, 🤖 Embodied and more.
• Easily extensible to new environments via a modular backend design.

Exploring Mechanisms of Agent Data Synthesis and Training

• Agentic Model Consolidation: Jointly and Stably train a unified model on mixed trajectories from all domains.

Innovative High-Value Agent Benchmarks

• Offer a suite of novel, high-quality benchmarks purpose-built for evaluating agentic capabilities.
• Designed to expose real-world challenges that existing benchmarks overlook, driving meaningful progress in agent research.

⚙️ Data Synthesis Method

AgentFlow synthesizes high-quality agent training data through a three-stage pipeline: Trajectory Sampling → Trajectory Selection → QA Synthesis.

1. Trajectory Sampling. An LLM-driven agent iteratively explores a sandbox environment starting from seed inputs. At each step it proposes a tool call, executes it, and records the observation, building a branching trajectory tree with concurrent expansion and action de-duplication.

2. Trajectory Selection. All root-to-leaf paths are scored by depth, information richness, and tool diversity, then selected with strategies, ensuring high-quality content.

3. QA Synthesis. For each selected path, the LLM generates a multi-hop, factoid QA pair grounded in the collected observations, with built-in quality checks.

📦 Installation

git clone https://github.com/OpenDCAI/AgentFlow
cd AgentFlow
bash install.sh # Install core dependencies

Optional dependencies:

bash install.sh --ml # + ML/DL (torch, transformers, etc.)
bash install.sh --cloud # + Alibaba Cloud SDK
bash install.sh --all # Install everything

All dependencies are listed in requirements.txt. See bash install.sh --help for details.

🛠️ QuickStart

We take WebAgent data synthesis as an example.

Step 1: Launch the sandbox with WebAgent sandbox config.

./sandbox-server.sh --config configs/sandbox-server/web_config.json \
--port 18890 \
--host 0.0.0.0

Step 2: Synthesize QA with WebAgent synthesis config.

fromsynthesisimportsynthesizesynthesize(config_path="configs/synthesis/web_config.json")

Step 3: Synthesize trajectory with WebAgent trajectory config.

fromrolloutimportpipelinepipeline(config_path="configs/trajectory/web_trajectory.json")

Step 4: After training the model, serve it with vLLM.

vllm serve \
--model YOUR_TRAINED_MODEL \
--served-model-name webagent \
--tensor-parallel-size 8 \
--enable-auto-tool-choice \
--tool-call-parser hermes \
--port 8222

Step 5: Infer on trained Agentic model with infer config.

fromrolloutimportpipelinepipeline(config_path="configs/infer/web_infer.json")

⚙️ Configuration

Purpose	Config Path
🖥️ Launching Sandbox	configs/sandbox-server/
🧪 Synthesizing QA	configs/synthesis/
🔄 Trajectory Rollout	configs/trajectory/
🚀 Model Inference	configs/infer/

🌟 AgentFlow Agent Family

Papers

AgentFlow also has an extensive agent family. You can find more information in the following paper:

[1] DocDancer: Towards Agentic Document-Grounded Information Seeking

[2] RAGShaper: Eliciting Sophisticated Agentic RAG Skills via Automated Data Synthesis

[3] Exploring Information Seeking Agent Consolidation

[4] BrowseComp-V3: A Visual, Vertical, and Verifiable Benchmark for Multimodal Browsing Agents

Models

Agent	🤗 HuggingFace
MM-Doc	DocDancer
RAG	RAGShaper
DeepResearch	DeepResearch Agent
General-datamix	Agent-datamix
General-RegMeanpp	Agent-RegMeanpp

Datasets

Agent	🤗 HuggingFace
MM-Doc	DocDancer
RAG	RAGShaper
DeepResearch	DeepResearch Agent

Benchmarks

BrowseComp-V3

A challenging benchmark of 300 hand-crafted multimodal questions for evaluating web browsing agents. It features deep multi-hop, cross-modal reasoning across diverse domains, with publicly searchable evidence and expert-validated subgoal-driven process evaluation. Even SOTA models like GPT-5.2 achieve only 36% accuracy. Includes OmniSeeker, a general multimodal browsing agent framework, along with full rollout and LLM-judge evaluation pipelines.

📄 Project Page · 🤗 Dataset · 💻 GitHub

🧪 Overall Performance

Qwen3-30B-A3B-Think

Level	Strategy	Web: GAIA (Acc.)	Web: BC (Acc.)	Web: BC-zh (Acc.)	Doc: MMBD (Acc.)	Doc: DocB (Acc.)	RAG: HotPotQA (EM/F1)	RAG: AmbigQA (EM/F1)	RAG: Bamboogle (EM/F1)
Data-level	Data Mixing	64.08	28.00	34.00	63.59	83.29	38.00 / 42.53	49.50 / 58.84	53.10 / 60.20
Parameter-Level	RegMean++	60.19	22.50	28.00	64.66	80.76	45.50 / 58.27	58.80 / 69.36	52.80 / 66.48

🔗 RAG Agent Case and Performance

Agentic RAG is an approach where an autonomous agent actively decides how and when to retrieve information and reason over it to accomplish a task.

Models	Bamboogle EM	Bamboogle F1	PopQA EM	PopQA F1	NQ EM	NQ F1	AmbigQA EM	AmbigQA F1	Avg EM	Avg F1
Prompt-Based Methods
IR-COT	16.0	27.9	32.4	39.9	19.3	35.5	24.5	40.6	23.1	36.0
RECOMP	21.7	28.6	40.5	45.8	–	–	–	–	–	–
Search-o1	30.4	39.9	47.0	50.0	30.3	40.7	42.5	53.4	37.6	46.0
Learning-Based Methods
Search-R1	30.4	43.2	41.3	46.4	36.0	45.0	49.2	60.4	39.2	48.8
ReasonRAG	22.4	29.1	41.1	44.4	28.1	38.9	39.7	51.9	32.8	41.1
HL-Data 4.5k	50.4	67.5	35.2	48.3	31.5	47.4	52.1	69.0	42.3	58.0
Ours
RAGShaper 4.5k	58.5	70.3	37.4	47.8	38.3	50.0	61.3	71.4	48.8	59.8
RAGShaper 6.5k	60.0	72.6	38.9	49.6	41.3	54.8	61.1	71.1	50.3	62.0

🙋 QuestionAmajorliteraryworkcommissionedbytheHolyRomanEmperorwhosereignbeganin1508waspartofhisgrandartisticlegacy. Whilethispatroncommissionedfamousmanuscriptanthologiesduringthisperiod, thisspecificallegoricalepicwasdistinctivelydesignedfortheprintingpresstoensureawideraudience.**Whatistheexactpublicationyearofitsfirstedition?**
💡 Answer1517

🔬 Document Agent Case and Performance

Document agent answers complex questions over multi-page documents by navigating, extracting, and reasoning across heterogeneous content—including text, tables, charts, and images.

Benchmark Results Comparison

Method	Model	MMLongBench-Doc acc	F1	LasJ	DocBench LasJ
**VLM
OCR-based Baseline
Tesseract	GPT-4o	30.1	30.5	—	—
Tesseract	Gemini-2.0-Flash	39.6	37.2	—	—
RAG-based Baseline
VisRAG	GPT-4o	29.0	27.8	—	—
RAGAnything	GPT-4o-mini	42.8	—	—	63.4
Prompt-based Agent
Doc-React	GPT-4o	38.1	38.3	—	—
MDocAgent	GPT-4o	42.0	—	—	—
SimpleDoc	Claude-4-Sonnet	—	—	58.6	—
DocLens	Claude-4-Sonnet	—	—	63.3	—
Ours
DocDancer	Qwen3-4B (ft)	48.4	49.2	59.4	79.8
DocDancer	Qwen3-30B-A3B (ft)	54.4	53.9	65.3	81.2
Human Baseline	—	65.8	66.0	—	81.2

🙋 QuestionWhatisthedifferenceinpercentage-pointincreasebetweentheoverallmeanscoreimprovementshowninthebarchartofpre-testversuspost-testscoresandtheimprovementfortheTICPrincipleconceptreportedinthepercentagestable?
💡 Answer14.92%

🖱️ Data Analysis Agent Case

Example input table (CSV preview, e.g. race_results.csv):

Rank	Driver	Team	Laps	Time / Retire	Grid	Points
1	sébastien bourdais	n / h / 1 racing	96	45:42.0	2	33
2	justin wilson	rsports	96	+ 3.9 secs	3	27
3	graham rahal	n / h / 1 racing	96	+ 6.6 secs	4	25
4	simon pagenaud	team australia	96	+ 24.8 secs	7	23
5	paul tracy	forsythe racing	96	+ 28.1 secs	14	22
...	...	...	...	...	...	...

🙋 QuestionWhichfeaturehasthehighestimportanceinpredicting'time / retired'accordingtotheRandomForestmodel?
💡 Answerlaps

🖱️ NL2SQL Agent Case

Table Structure

actor Table (200 rows)

actor_id	first_name	last_name	last_update
1	PENELOPE	GUINESS	2026-02-05 16:18:42
2	NICK	WAHLBERG	2026-02-05 16:18:42
3	ED	CHASE	2026-02-05 16:18:42

film_actor Table (5,462 rows)

actor_id	film_id	last_update
1	1	2026-02-05 16:18:45
1	23	2026-02-05 16:18:45
1	25	2026-02-05 16:18:45

film Table (1,000 rows)

film_id	title	description	release_year	rental_rate	length	rating
1	ACADEMY DINOSAUR	A Epic Drama of a Feminist...	2006	0.99	86	PG
2	ACE GOLDFINGER	A Astounding Epistle of a...	2006	4.99	48	G
3	ADAPTATION HOLES	A Astounding Reflection of...	2006	2.99	50	NC-17

film_category Table (1,000 rows)

film_id	category_id	last_update
1	6	2026-02-05 16:18:48
2	11	2026-02-05 16:18:48
3	6	2026-02-05 16:18:48

category Table (16 rows)

category_id	name	last_update
1	Action	2026-02-05 16:18:42
2	Animation	2026-02-05 16:18:42
3	Children	2026-02-05 16:18:42

inventory Table (4,581 rows)

inventory_id	film_id	store_id	last_update
1	1	1	2026-02-05 16:18:42
2	1	1	2026-02-05 16:18:42
3	1	1	2026-02-05 16:18:42

rental Table (16,044 rows)

rental_id	rental_date	inventory_id	customer_id	return_date	staff_id
1	2005-05-24 22:53:30	367	130	2005-05-26 22:04:30	1
2	2005-05-24 22:54:33	1525	459	2005-05-28 19:40:33	1
3	2005-05-24 23:03:39	1711	408	2005-06-01 22:12:39	1

Synthesized Question

Which film categories have more than 10 actors who have appeared in popular films, and how many such actors are there in each category?

Answer SQL

WITH film_popularity AS (
SELECTf.film_id, COUNT(r.rental_id) AS rental_count FROM film f JOIN inventory i ONf.film_id=i.film_idJOIN rental r ONi.inventory_id=r.inventory_idGROUP BYf.film_id
), top_popular_films AS (
SELECT film_id FROM film_popularity WHERE rental_count > (SELECTAVG(rental_count) FROM film_popularity)
), actor_film_count AS (
SELECTa.actor_id, COUNT(f.film_id) AS film_count FROM actor a JOIN film_actor fa ONa.actor_id=fa.actor_idJOIN film f ONfa.film_id=f.film_idGROUP BYa.actor_id
), top_actors AS (
SELECT actor_id FROM actor_film_count WHERE film_count > (SELECTAVG(film_count) FROM actor_film_count)
), actor_category_data AS (
SELECTfa.actor_id, c.nameAS category_name FROM film_actor fa JOIN film_category fc ONfa.film_id=fc.film_idJOIN category c ONfc.category_id=c.category_id
) SELECTc.nameAS category_name, COUNT(a.actor_id) AS actor_count FROM top_actors ta JOIN actor_category_data acd ONta.actor_id=acd.actor_idJOIN category c ONacd.category_name=c.nameGROUP BYc.nameHAVINGCOUNT(a.actor_id) >10ORDER BY actor_count DESC;

🖱️ GUI Agent Case

GUI Agent Case

GUI_case.mp4

🙋 InstructionIwanttoauditallcommandaliasesonthisUbuntumachine, sopleaselaunchtheterminalfromtheGUI, identifyanyhomedirectoryconfigfilesrelatedtoshellstartup, andthengenerateaclean, sortedlistthatcombinesbothcurrentlyactivealiasesandthosehiddeninyourconfigurationfilessoIcanseethefulldefinitionsofcommandslikealertorll.

🖱️ Embodied Agent Case

Place the mouse on the yellow pad	Open the laptop
Place the cup on the blue box	Store the car in the basket

📜 License

Apache 2.0

✍️ Contributors

Role	Members
🎯 Project Leader	Zhengwei Tao (tttzw@pku.edu.cn), Jialong Wu (wujialongml@gmail.com)
🌟 Core Contributor	Bo Li, Guochen Yan, Qintong Zhang, Huanyao Zhang
💡 Contributor	Xinjie Lv, Haishan Lu, Yuan Xu, Haoyang Yao, Xingdi Ding
📣 Advisor	Kuan Li (UniPat.ai)
🏫 Supervisor	Wentao Zhang, Bin Cui

🤝 Community & Support

💬 Join the AgentFlow open-source community to ask questions, share ideas, and collaborate with other developers!

_{👆 Scan to join the community WeChat group 🎉}

🌍 Citation

If you use AgentFlow in your research, please cite:

@misc{omniagentsynth2026,
title={AgentFlow: Unified Agent Data Synthesis Framework},
author={AgentFlow Team},
year={2026},
howpublished={\url{https://github.com/OpenDCAI/AgentFlow}}
}