Skip to content
Ian Tharp edited this page Jun 4, 2026 · 2 revisions

Dynamic Subgraphs

A governed runtime for LLM-generated workflows. A stable Supervisor turns a prompt into a validated plan — data, never executable code — then compiles and runs it as a transient LangGraph workflow over a bounded, allowlisted vocabulary of node kinds. The hard problem it answers:

How do you let an agent design its own workflow at runtime without letting it generate and run arbitrary code?

The model proposes; a validator decides whether the proposal is safe; the runtime executes it and records every run as a replayable, diffable artifact.

pip install dynamic-subgraphs        # or: uv add dynamic-subgraphs
from dynamic_subgraphs import DynamicSubgraphs, EngineConfig, Model

engine = DynamicSubgraphs(EngineConfig(model=Model("openai", "gpt-5.4-nano")))
result = engine.run("Compare two sources on X and recommend one.")
print(result.response)

Start here

  • FAQ — "is this just a LangGraph wrapper?", "when should I not use it?", and other honest answers.
  • Concepts — the governed-runtime model, plans-not-code, the node vocabulary, nested subgraphs.
  • Guides — choosing a model (cloud / local / hybrid), recording & replay, troubleshooting.
  • Roadmap — where the project is going (eval layer, host-owned governance, extension seam).
  • Design Decisions — the red-teamed design specs behind the big changes.

Canonical references (in the repo)

The wiki is the living, narrative companion to the code-versioned docs in the repo. For anything that must track a specific release (API reference, recipes), the in-repo docs/ are authoritative.

Clone this wiki locally