**TL;DR:** For new apps the docs recommend event streaming via stream_events with version v3, which returns typed projections you consume independently. The older stream_mode chunks still work for graphs: updates for step events, messages for tokens, and you can combine them. Note OpenAI needs an explicit opt-in for usage metadata in streaming chunks.

## The fix

1. `stream_mode="updates"` emits an event after every agent step; `stream_mode="messages"` streams LLM token chunks; pass a list to get both.

2. Always pass a `thread_id` via config when streaming a conversation. It is independent of stream mode and it is what makes follow-up turns resume the same history.

3. Wrapping a create_agent as a node in a parent StateGraph? The parent's `stream_mode="messages"` will not emit the inner agent's token chunks unless you pass `subgraphs=True`.

4. `stream_mode="custom"` is for your own `writer()` events inside nodes.

5. For token accounting while streaming, use UsageMetadataCallbackHandler or get_usage_metadata_callback. Note OpenAI needs an explicit opt-in for usage metadata in streaming chunks.

## When to use this

- This covers exactly what the title says: LangChain streaming.
- You are setting this up for the first time, or auditing an existing setup.
- You want the key gotchas in one place before you start.

## When not to use this

- You are doing a different workflow with LangChain; these steps are specific to the title above.
- You need the full reference docs; this is the short path, not the manual.