Your Agent Streams Text But Breaks on Tool Calls. Here's the Fix.

Streaming tokens from an LLM is easy. You get a callback per token, you push it to the client, done. Then you add tool calls. The LLM starts streaming a tool input JSON character by character. You need to execute the tool (blocking, could take 3 seconds). Then you resume streaming. Meanwhile, the client is sitting there wondering if the connection dropped. Then you add multi-agent pipelines. Agent A streams into Agent B streams into Agent C. Which events does the UI show? All of them? Just the final output? Then a user's browser tab goes to sleep and they miss 40% of the stream. They refresh. Do they start over or resume? These are the failure modes that hit production streaming agents. Here's how to handle all of them. Start With the Event Envelope Don't pipe raw LLM tokens to your client. Normalize everything to a typed event: class EventType ( str , Enum ): TEXT_DELTA = " text_delta " TEXT_DONE = " text_done " TOOL_CALL_START = " tool_call_start " TOOL_CALL_INPUT = " tool_call_input