Claude chooses when to speak; Microsoft moves agent controls into the request path
Anthropic gave Claude full-channel context in Slack, Microsoft published run-time checkpoints, and OpenAI's own figures show a wide adoption gap between employees and subscribers.
Artificial Intelligence··Evening
Claude can read the full channel
Anthropic removed the classifier that judged Slack messages one at a time in its Claude Tag update. Claude now reads the full channel context and chooses one of four actions: answer inline, move the topic into a thread, route the message into an existing workstream or stay quiet. Its reach is limited to the intersection of what the agent can see and what the requesting person can see. That boundary is intended to prevent information leaking between channels. Anthropic reports an improvement of about 30 percent in deciding when to speak and when to stay quiet. Administrators can also set budget caps tied to agent identities.[1]
Microsoft enforces controls at run time
Microsoft's published architecture turns written AI policy into controls applied while a request is being processed. Microsoft Foundry's agent gateway handles authentication, token limits, quotas and policy enforcement. It provides central authentication, rate limits, address restrictions and audit logging for Model Context Protocol tools. The Agent Control Specification defines checkpoints across agent inputs, model calls, tool execution and outputs. The architecture covers 9 domains: policy, data, model, observability, evaluation, security, identity and access, audit, and agent governance. Microsoft also released an open-source Agent Governance Toolkit.[2]
OpenAI's internal use has not carried over to subscribers
In June 2026, 98 percent of OpenAI employees were using its coding agent Codex. The share was 17 percent among organisational subscribers and below 1 percent among individual subscribers. Last month the company moved ChatGPT Work, its adaptation of Codex for non-engineers, to the lowest subscription tier at 20 dollars a month. The desktop and mobile app has 20 million users, compared with more than 1 billion online users for ChatGPT. OpenAI's product team says the system can carry complex tasks from start to finish. Desktop lead engineer Andrew Ambrosino acknowledges that it may pull from a private message while writing a document.[3]
