Ruhao Liu‡ Qi Li‡ Shizun Wang§ Xinchao Wang†
xML-Lab, National University of Singapore
‡Equal contribution §Project Lead †Corresponding Author
English | 简体中文
DynaFlow supports direct model reasoning and model-generated agent workflows.
| Mode | Behavior | Configuration |
|---|---|---|
direct |
The main model answers without tools. | direct.json |
dynamic |
The main model chooses a direct answer or a Workflow. | dynamic.json |
forced |
The main model must complete a Workflow with successful subagent delivery. | forced.json |
- Python 3.11+.
- For
dynamicandforced: Node.js 24+, npm, and Docker with Linux containers. - An OpenAI-compatible streaming Chat Completions API that returns finish reasons and usage. Workflow models must support tool calls.
Run all commands from the project root.
Copy .env.example to .env and fill in your API URL, model name, and key:
DYNAFLOW_BASE_URL=https://your-api-host/v1
DYNAFLOW_MODEL=your-model-name
DYNAFLOW_MODEL_ADAPTER=openai
DYNAFLOW_API_KEY=Use adapter deepseek for DeepSeek. The CLI loads .env automatically; existing environment variables take precedence.
For dynamic and forced, also configure the web search and page retrieval credentials described in .env.example. These services are checked before the demo starts, even if dynamic later answers directly. Subagents and page summaries reuse the main model API by default.
Required for dynamic and forced; skip this step for direct. Start Docker with Linux containers, then run:
docker info
npm --prefix workflow_runtime ci --ignore-scripts
npm --prefix workflow_runtime run build
docker pull node:24.14.0-bookworm-slim
docker build --file tool_runtime/Dockerfile --tag dynaflow-tools:py311-v1 .Docker installs the Workflow tools' Python calculation libraries from requirements.txt. The host program uses the Python standard library for all three modes; no additional pip install is needed on your machine.
On Windows, use npm.cmd if PowerShell blocks npm, and keep checkout and output paths short.
Choose a mode:
python scripts/dynaflow.py demo --mode direct
python scripts/dynaflow.py demo --mode dynamic
python scripts/dynaflow.py demo --mode forcedEach command runs the two questions in examples/demo/questions.json using real APIs. No external dataset is needed. A successful run means all questions completed with valid answer delivery; answer correctness is not scored.
Results are saved under runs/demo-<mode>/:
| Path | Contents |
|---|---|
results/ |
Model answers |
calls/ |
API call records |
comparison/ |
Run summary, token usage totals, and answers in CSV format |
workflows/, tools/ |
Workflow and tool execution records, when used |
To run again, select a new directory with --out runs/my-demo.
If you find our work or this codebase helpful, please cite our paper:
@article{liu2026dynaflow,
title={DynaFlow: Learning Free-Form Multi-Agent Orchestration},
author={Liu, Ruhao and Li, Qi and Wang, Shizun and Wang, Xinchao},
journal={arXiv preprint arXiv:2610.34467},
year={2026}
}DynaFlow's original code is licensed under the MIT License.
The workflow engine in workflow_runtime/src retains its original MIT license and copyright notice.
