Starting November 11, Google Workspace MCP servers will reject any OAuth client that is not marked as an AI agent. Agentic OAuth, a single checkbox in the Google Cloud console, is now the admission ticket for agents that want to reach Gmail, Drive and Calendar through the Model Context Protocol. The change looks administrative. It is the first time a major workplace platform has required agent traffic to identify itself at the credential layer before it can touch company knowledge.
What Google changed
Google is rolling out Workspace MCP servers that expose Gmail, Drive, Calendar and other apps to external AI agents. Ahead of that launch, Google told developers by email that "only OAuth clients marked as agentic will be able to call Workspace MCP servers", as reported by Apps Script Pulse. The setting is a flag on the OAuth client, named is_agentic, switched on by selecting "Use this client for an AI-powered agent".
The dates are tight. Developers have been able to edit existing clients since October 12. On November 11, Workspace MCP servers begin rejecting unmarked clients. On November 12, the setting becomes permanent on existing clients: miss the window and the only fix is a new OAuth client ID, with new consent screens and every connected tool reconfigured, as the builder checklists now circulating point out.
Google's own documentation sets two further rules. An application that serves both people and agents must maintain separate OAuth client IDs for each: a standard client for human sessions, an agentic client used exclusively for AI agents, automated tool execution and MCP integrations. And the flag grants nothing by itself. Tagging a client does not add scopes or privileges, and agentic clients remain subject to user consent and OAuth app verification.
Why classification comes first
Until now, an agent reading a Drive folder and a person opening the same folder could arrive through the same credential, indistinguishable in the log. Governance cannot start there. As one daily briefing on the change put it on October 13, this is the protocol-level version of a simple rule: you cannot govern agent traffic until you can tell it apart from everything else. Google makes the same case in operational terms: separate client IDs give clean monitoring, and let administrators apply dedicated access and security policies to agents without disrupting human access to the standard application.
Read alongside last week's post on this blog, "Agent Identity: Microsoft Gave Agents Their Own Accounts. The Memory Is Shared.", the pattern is clear. Microsoft is giving its own agents identities inside the suite. Google is making outside agents declare themselves at the door. Both platforms have decided that anonymous agent access to company knowledge is over.
What agentic OAuth does not do
Classification is a label, and Google is candid that it is a self-declaration: the checkbox identifies a client as representing agents that act or retrieve context on behalf of users. It says nothing about behaviour. Within the scopes a user has consented to, the flag does not decide which documents an agent should retrieve, whether an answer gets checked before it is used, or whether what an agent writes back should enter the company record at all.
Those decisions sit one layer up, in the product the agent connects to. That is the layer this site's pillar article, "How Jilu Works", is about: an agent that answers only with the permissions of the person who asks, that has every action logged against that person, and whose drafts wait for a person to review and publish them before they join the shared record. A platform flag can tell an administrator that traffic is agentic. It cannot tell them whether the knowledge that traffic produced was ever reviewed.
The edges are already showing. On October 13, a developer asked Google how to classify an Apps Script project whose automatically generated OAuth client cannot be modified at all. Classification schemes leak wherever credentials were created before the scheme existed, which in most companies is everywhere.
The deadline is the small part
For builders, November 11 is a checklist item: inventory your OAuth clients, split human and agent traffic, tick the box before it locks. The larger shift is the question administrators can now ask in a precise form. Not whether an agent reached company knowledge, but under which declared identity, with whose permissions, and into what record its work went. Agentic OAuth answers the first part of that question. The rest is the company brain problem, and a checkbox does not solve it.