Anthropic’s Claude Tag, its Slack-based AI agent, now reads full conversation context instead of evaluating messages one at a time. The change improves its ability to decide when to jump into discussions unprompted by about 30%, according to the company. Anthropic frames this as a move from AI as a personal tool to an organizational colleague, citing three technical pillars: the Model Context Protocol for data access, a threshold of model intelligence enabling proactivity, and a Slack integration that embeds Claude where teams already work.
The update lets Claude choose among four actions—reply inline, start a thread, route the message, or stay silent—using channel context, memory, and standing instructions. Anthropic built restraint into the system, making the agent dormant in channels where it repeatedly adds no value. The company argues collaborative AI outperforms solo tools by reducing handoffs, pointing to internal use where engineers now spend more time on strategy and less on waiting for data analysis.
Security remains a focus, with Anthropic layering defenses against prompt injection attacks, including model-level classifiers, compliance APIs, and data loss prevention integrations. Permissions are restricted to the most limited intersection of what the agent can see and what the user can access, preventing cross-channel data leaks.
Pricing for expanded context remains unsettled, with Anthropic treating it as part of an experimental phase. The company is co-designing deployments with customers but warns enterprises to assume eventual billing for context-heavy usage.
