Comparison: LangGraph, n8n, Google ADK
Know the other circles before drawing your own.
Sanctum does not try to out-feature the incumbents. It competes in one niche — local-first agent orchestration you can audit whole — and reimplements the ideas that matter (cycles, reducers, checkpoints, streaming, middleware) in a core small enough to read in an afternoon.
| Sanctum | LangGraph | n8n | Google ADK | |
|---|---|---|---|---|
| Execution model | Cyclic state graph, BSP supersteps | Cyclic state graph, Pregel-inspired | Visual workflow automation (mostly linear/branching) | Agent framework with delegation trees |
| Core dependencies | None (Python stdlib) | LangChain ecosystem | Node.js platform | Google Cloud SDK stack |
| LLM coupling | Local-first Oracle interface: Ollama, llama.cpp, vLLM, LM Studio, in-process GGUF |
Any provider, best with LangChain integrations | Nodes per provider | Gemini-first |
| Small-model tool-calling | First-class: tolerant JSON repair, correction rounds, prompted fallback | Provider-dependent | Provider-dependent | Gemini-native |
| State merging | Per-channel reducers, deterministic insertion order | Per-channel reducers | Node outputs | Session state |
| Persistence | JSON Seals: memory / SQLite / Postgres; append-only history | Checkpointers: memory / SQLite / Postgres | Built-in DB | Sessions/Memory services |
| Human-in-the-loop | interrupt() + resume with injected updates |
interrupt() / breakpoints |
Manual approval nodes | Callbacks |
| Time-travel | From any Seal, honest append-only replays | From any checkpoint | Re-run executions | — |
| Streaming | Typed Omens, combinable modes, Sigil writer |
Multiple stream modes | UI executions | Event streams |
| Middleware | Ward pipeline (transform/veto deltas, observe omens) | — (callbacks/hooks) | — | Plugins/callbacks |
| Observability | Local: JSONL audit, usage tally, single-file HTML trace viewer — no external services | LangSmith (SaaS) or callbacks | Built-in UI | Cloud Trace |
| Resilience | Per-Sigil timeout/retries/backoff/fallback policies | Node retry policies | Node retries | Runner config |
| Maturity & ecosystem | Young, small, thesis-grade docs | Production-grade, large ecosystem | Mature product, huge integration catalog | Backed by Google, growing |
Choose…
- Sanctum when you run local models, need to understand and audit every superstep, want observability without SaaS, or are studying how agent engines work (the architecture document is the design rationale, trade-offs included).
- LangGraph when you need the batteries: hundreds of integrations, a managed platform, and production mileage.
- n8n when the job is automating SaaS workflows visually, not building agent loops.
- ADK when you build on Google's stack and Gemini.