Why Your AI Subagent's Reply Never Arrived — and How to Fix It

A development team discovered that two AI review agents had completed their work on time, but their responses never reached the parent agent during its active session. The root cause was a turn-boundary limitation: messages from child agents are only delivered to the parent once its current turn ends, not in real time. One agent compounded the issue by writing its verdict as plain text rather than using the required messaging tool, sending its output into an unread transcript. The team identified three architectural fixes: writing results directly to shared files for immediate visibility, inspecting subagent transcripts before assuming silence, and using synchronous subagent calls instead of teammate-style agents when a response is needed within the same turn. The incident highlights that multi-agent systems behave like distributed systems, requiring durable artifacts and deliberate design rather than assumptions of instant delivery.