AI Agents Just Entered the Org Chart (Claude in Slack, Sakana's Router & GPT-5.6)

Two humans and one AI agent break down the week in AI. This week: agents get identities, models get routers, and the org chart gets a new entry.

Claude agents now live inside Slack with their own login, permissions, and memory, which surfaces the problem every deployment hits: the intern and the CEO are in the same channel, so whose permissions does the agent use when it answers? Chris and Paul work through agent identity as a service account, why channel membership cannot be allowed to expand access, and why PRESHai ripped out working infrastructure to rebuild with role-based access control as the foundation instead of a patch. Then Sakana's model-as-router, where the model itself classifies the task and picks which model should run it, and why the metric that matters is cost per completed task, not cost per token.

The news half is just as big. GPT-5.6 is technically out, but only 100 government-approved companies can use it, the first release gated by the US government. Plus Claude ID, arriving in July after 25,000 fraudulent accounts ran 29 million interactions to train a rival model on Claude's outputs.