Anthropic shipped an update to Claude Tag, its persistent AI teammate that lives inside Slack channels, changing how the system decides when to proactively contribute to a conversation without being directly mentioned. Previously, a lightweight classifier evaluated each incoming message in isolation, making a single yes-or-no call about whether to respond based on that one message alone. The problem with this approach is that plenty of genuinely useful moments span multiple messages from different people, none of which individually reads as a request for help. Anthropic's example is two engineers separately posting a bug theory and the corresponding evidence in the same channel, with neither message directly addressed to Claude or to each other in an obvious way. Under the old system, Claude would correctly conclude there was nothing to do on each message read alone, and miss the connection entirely. The new version gives Claude visibility into the full channel context, plus its own memory and any standing instructions a team has configured, so it can choose from four behaviors: reply inline for short, verifiable questions, open a thread for deeper investigation, route a new message into work it already has running, or simply stay silent. Anthropic reports this makes Claude roughly 30 percent more accurate at judging when to speak up versus stay quiet, a meaningful problem for any always-on workplace agent, since an assistant that interrupts too often gets muted or removed, while one that says too little gets ignored. The update also means Claude acknowledges a relevant conversation within seconds rather than only after being explicitly asked, and it comes at no extra cost within Claude Tag's existing Team and Enterprise plans. For anyone designing agentic tools that sit passively in a communication channel waiting to be useful, this is a concrete illustration of why context window scope, not just model capability, determines whether an agent feels helpful or annoying.