Separate tracks were the speaker attribution. The lock is what takes them away — not by failing, but by quietly leaving one stream unrecognised.
An undocumented platform limit — Apple's SFSpeechRecognizer silently kills a second concurrent session with no error — took an AI research agent two minutes to find and explain. A human developer typically discovers it only by hitting the wall and searching forums for half a day. That's the real leverage of agentic engineering: research speed on platform behavior nobody wrote down, not code generation.