⚙️ Dynamic Workflow Agents

Build a multi-agent runtime from scratch. Give it a real job. Measure what it actually improves.

Explore the notebooks · How it works · Results · Get started · Repository guide

---

This repository contains two self-contained editions of **Dynamic Workflows from Scratch**. In each notebook, a lead agent writes a phased plan, a plain-Python runtime executes it with parallel workers, and experiments test where that design helps. The main job is a planted-bug audit of a large algorithm library. The same engine later grades maths homework and groups customer messages.

🔍 Benchmark	🧪 Experiment	⚙️ Runtime	📊 Deliverable
72 planted bugs	Solo, fixed, and lead-written plans	Parallel phases, events, threads, budgets, and resume	2 executed notebooks with outputs and figures

[!IMPORTANT] The Claude edition completed a live run, offline replay, and prose audit. The Qwen edition completed a live run and passes its notebook checks, but its final offline replay and line-by-line prose audit are still pending. Qwen figures below come from saved live experiment records; treat their interpretation as provisional. Smoke-run numbers are scripted and are never evidence of model performance.

Explore the notebooks

	🔷 Claude edition	🔶 Qwen edition
Open the notebook	Read `claude/dynamic_workflows.ipynb`	Read `qwen/dynamic_workflows.ipynb`
Provider in the recorded run	Anthropic API	ModelScope's Qwen endpoint
Accounting	Per token, in cents	Per call, in Magicubes
Notebook spending cap	2,500 cents