LangGraph

Among OSS agent frameworks, LangGraph is one of the closest analogues to propose → gate → execute: interrupt() is the seam Headroom and similar HITL layers hang on. Use it when durable graph state and visible branch/repair paths matter more than a simple ReAct loop. Keep it on the observe/propose side of an agentic OS — useful drafts and tool plans, with policy authority still outside the graph. Prefer Temporal / Inngest when workflows must span services and languages; prefer plain LangChain agents when you do not need checkpoints or interrupts.

1. Declare graph state (often message-centric) and nodes (LLM, tools, deterministic functions).
2. Compile with a checkpointer for thread-scoped short-term memory; optionally a store for cross-thread facts.
3. Invoke or stream; on interrupt, a human or policy reviews state, then resume.
4. Only then attach irreversible side effects behind an explicit gate (not inside an unchecked node).

Related: langchain · langmem · langsmith · temporal · headroom · topics/12-central-nervous-system

• Use when: you need durable multi-step agent graphs with checkpoints, branching, and HITL pauses in Python (or LangGraph.js).
• Skip when: a single-shot chain or prebuilt LangChain agent is enough, or when cross-language durable workflows belong in Temporal/Inngest.
• Prefer instead: temporal for polyglot mission-critical workflows; langchain for model/tool loops without graph persistence; headroom when the hard problem is escalation calibration, not graph runtime.

• No built-in Brier/calibration — interrupt ≠ good policy; measure thresholds elsewhere (e.g. Headroom / T16).
• Ops: checkpointer and deployment choices dominate failure modes (lost threads, stuck interrupts).
• Scope: orchestration only — not the ledger, Case Graph, or approval authority.
• Ecosystem gravity: docs and examples lean LangChain/LangSmith; you can use LangGraph without LangChain models, but most packs assume the stack.