1.⁠ ⁠We kept hitting a "Stream closed" error when our agent sent WhatsApp messages. It turned out to be a bug in the Claude Agent SDK’s MCP layer, not our code. When the agent called multiple tools quickly (like sending several messages in one turn), the SDK sometimes closed the stream too early. It got worse when users sent new messages while the agent was still running, so multiple runs overlapped. To make this reliable at scale, we’d add per-user locks so only one agent run per user runs at a time, and we’d reduce how many tool calls we make per turn.

2.⁠ ⁠The biggest headache was getting the agent to handle video generation without crashing. Videos take forever to make - like 1-2 minutes for the AI to generate voiceover, find clips, and stitch everything together. But the agent session times out way before that.
So what happened? Agent starts making the video, session dies, video generation keeps running somewhere in the background, but the agent never knows it finished. User gets nothing on WhatsApp even though the video actually got made.
Spent way too long debugging why videos were generating but not sending. Logs showed everything working but users saw nothing.
Fixed it by splitting things up - agent just acknowledges the request immediately and kicks off video generation as a background task. That task runs separately, does all the heavy work, and sends the video directly to WhatsApp when done. No more timeout issues.