Weekly AI Tools

In two weeks the agent acquired a trigger it does not control, a credentialled session it can reuse, and a schedule nobody attends. Then OpenAI pointed the same architecture at the workspace itself and, unlike everyone else, wrote down what it may not do.

In eight days Anthropic shipped scanning for third-party skills, Google shipped an administrator approval queue for them, and Treasure AI redefined a saved workflow as a skill. Software supply chain vocabulary, applied to prose. Meanwhile the read-only phase of agent integration ended without an announcement.

Three agent platforms published releases in the same week and shipped no new model, no new agent type and no new autonomous behaviour. What they shipped instead was enterprise identity, tool-call tracing and consumption billing, which are the three things that were actually blocking rollouts.

One frontier model landed this week, and it was not the story. The consequential releases were an administrator key that makes MCP tool approvals expire with the session, and a console for finding out which agents you are already running.

The first edition of a new weekly series on agentic work platforms. This week the big three stopped selling capability and started selling control: OpenAI Presence, cryptographic agent identity in Gemini Enterprise, and governed publishing in Microsoft Copilot.