HeyHomees Lab Notes · No. 10 · National University of Singapore · Published October 27, 2026
By HeyHomees Newsroom ·
In 30 seconds
- The main model creates free-form workflows that direct fixed sub-agents for tasks
- In Forced mode DynaFlow scores 88.08% overall accuracy across six benchmarks
- Running DynaFlow needs Python 3.11+, Node.js 24+ and Docker with Linux containers
What is it?
DynaFlow is a framework that integrates free-form executable workflow synthesis into the action space of the main agent, letting it decide when to collaborate and generate task-specific programs that coordinate sub-agents
How does it work?
The main agent is trained via supervised fine-tuning on distilled trajectories while all sub-agents stay fixed; workflows can run sub-agents sequentially or in parallel and include verification steps, with execution traces feeding back into later actions
How well does it do?
Across six benchmarks DynaFlow attains 88.08% overall accuracy in Forced mode, beating the strongest baseline by 10.90 percentage points using the same Qwen3.5-27B backbone
What does it need to run?
It requires Python 3.11+, Node.js 24+, npm, and Docker with Linux containers; the code is released under the MIT License
By the numbers
- Overall accuracy in Forced mode: 88.08%
- Improvement over strongest baseline: 10.90 percentage points
- Number of benchmarks evaluated: six
- Required Node.js version: 24+
The honest read
The reported 88.08% accuracy and 10.90-point gain are self-reported results on six internal benchmarks and compare only to baselines evaluated with the same Qwen3.5-27B backbone; the paper does not provide external or real-world deployment metrics. All sub-agents are fixed, so the numbers reflect orchestration learning rather than improvements to the underlying executors. The code runs under the MIT License but performance depends on external APIs and Docker setup as described. The figures do not address other modes (direct, dynamic) or runtime speed, and no download size or model parameter count is given.
What we don't know yet
- Exact model parameter count or download size
- Performance of the direct and dynamic modes
- Latency or compute cost of running the workflows
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.