docs(graph): llm nodes stream internally; node timeout bounds the chat loop, read_timeout only stalls

Graph-Agents.md

@@ -743,6 +743,18 @@ Unlike the graph-level `settings.timeout`, a node timeout still gives the
 graph a chance to recover or report through its `fallback` route.
 `timeout: 0` (or unset) disables the bound.
 
+`llm` nodes stream the completion internally, quietly: chunks are
+accumulated as they arrive and nothing is rendered. This matters for the
+interplay with the client-level `read_timeout` ([Clients](Clients),
+default 300s), which fires between successive reads. Because the node
+receives bytes throughout generation, `read_timeout` only bounds a genuine
+stall between chunks, while the node `timeout` bounds the node's chat loop
+(the structured-output extraction pass that follows it is a separate call
+outside the node `timeout`, bounded only by `read_timeout` stalls). A node `timeout` above 300s therefore
+genuinely allows generations longer than 300s; without streaming, the first
+byte would not arrive until the reply was complete and `read_timeout` would
+cap it regardless.
+
 ### Retries (`max_attempts`)
 
 `max_attempts` retries the LLM call **only on transient errors**. The