Repository navigation
FAQ
No. LangGraph is the execution engine, but it stays behind a compiler/ and runtime/ boundary. What Dynamic Subgraphs adds is governance: the planner emits a declarative GraphSpec (data, not code), a deterministic validator decides whether that spec is acceptable (topology, allowlists, budgets, reachability, cycles, input provenance) before anything compiles, and every run is recorded as a replayable artifact. With plain LangGraph you hand-author the graph; here the model proposes one and the system decides whether to trust it.
Reach for something simpler when:
- The shape is known. If you can draw the DAG ahead of time, hand-author a fixed LangGraph graph — cheaper and more predictable.
- The task is small. A frontier model in a plain tool loop already decomposes a 1–3 step task in-context; a planning round-trip is overhead.
- Your hard problem is global consistency, not orchestration. Isolated child envelopes are great for blast-radius but work against shared canonical state — pair DS with a retrieval/consistency layer instead.
DS earns its keep when the workflow shape varies per input and you need the run governed and auditable.
Three things a ReAct loop can't give you for free: (1) the model proposes the workflow but never executes arbitrary code; (2) its capability surface is an allowlisted, bounded vocabulary; (3) recursion is depth- and budget-capped, and every run — success or failure — is inspectable and replayable. That audit trail is the point.
Those are multi-agent / persona frameworks. DS is an orchestration substrate — the layer that decides what shape a piece of work should take and runs it under governance. It's complementary, not competing: you could build a multi-agent system on top of a governed runtime.
No — recording is opt-in. By default the engine runs in-memory and writes no files, so embedding it never clutters your tree. Turn on a Recording policy (presets or per-artifact) when you want the spec / trace / Mermaid diagram / cost captured under runs/<run_id>/. See Guides.
Yes. Bring your own model/endpoint: Model.lmstudio(...), Model.ollama(...), or any OpenAI-compatible server via Model.openai_compatible(...). One caveat — the planner needs to be a capable model (small local models often emit invalid plans); a good pattern is a capable cloud planner with cheap/local workers. See Guides.
It's an honest alpha — typed (py.typed), tested across OSes and Python 3.11–3.13, and Apache-2.0. The governance contract is being hardened toward "impossible to bypass" (see Roadmap); until that lands, treat it as a strong alpha, not production-safe.
SemVer, and while 0.x a breaking change bumps the minor, a feature/fix bumps the patch. The full policy (what counts as public, deprecation, etc.) is in docs/api-stability.md.