From 00c01f4c9fa7cc266351a428a1da320f036fa998 Mon Sep 17 00:00:00 2001 From: Brian Love Date: Thu, 27 Aug 2026 19:58:31 -0700 Subject: [PATCH] docs(langgraph): state isolation in the matrix is designed, not automatic MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The Multi-Agent Supervisor sample directly above the decision matrix adds compiled children as plain nodes on a shared OrchestratorState, then the matrix claimed that pattern gives you 'Isolated per subagent'. It doesn't — parent and child read and write the same channels. Change the cell to 'Only if you design it' and state the mechanism inline, since the subgraphs guide also uses a shared MessagesState and so can't serve as the pointer. Co-Authored-By: Claude Opus 5 --- .../content/docs/langgraph/concepts/agent-architecture.mdx | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/apps/website/content/docs/langgraph/concepts/agent-architecture.mdx b/apps/website/content/docs/langgraph/concepts/agent-architecture.mdx index e199170ea..0c83a294b 100644 --- a/apps/website/content/docs/langgraph/concepts/agent-architecture.mdx +++ b/apps/website/content/docs/langgraph/concepts/agent-architecture.mdx @@ -686,9 +686,11 @@ builder.add_conditional_edges("supervisor", route_to_agent) | Tool count | 1-10 | 1-10 | 10+ across specialists | | Task complexity | Single domain | Single domain, high stakes | Cross-domain | | Latency budget | Low | Medium (human wait) | Higher (multiple LLM calls) | -| State isolation | Shared | Shared + interrupt | Isolated per subagent | +| State isolation | Shared | Shared + interrupt | Only if you design it | | Angular complexity | Low | Medium | Higher | +State isolation is the one row that is not automatic. Adding a compiled graph as a node — as in the supervisor example above — runs the child against the parent's state schema, so parent and child read and write the same channels. To actually isolate a specialist, give its `StateGraph` its own state schema and share only the keys you want crossing the boundary: LangGraph passes a subgraph node through the keys the two schemas have in common. + Begin with a single agent and tools. Add human-in-the-loop when you need approval flows. Graduate to multi-agent only when a single agent's context window cannot hold all the tools and instructions it needs.