Your Newest Coworker Has an Email Address

On October 8, Google Cloud announced a universal Gemini agent for work, adapted from CEO Thomas Kurian’s keynote at Gemini at Work 2026. The agent answers questions, writes and runs code, and works inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, from any device and any channel, carrying the same memory, skills, and controls with it everywhere. Industry-specific agents for financial services and legal are already in preview; government, healthcare, and retail are coming. The release note is explicit: it learns from each interaction with the user.
Read that again. An agent with an email address, a calendar, a Drive folder, and a memory that persists across sessions. Not a chatbot you visit. A named coworker that sits inside your systems of record and keeps getting smarter about them.
This is the moment the identity problem stops being theoretical. SailPoint’s Horizons of Identity Security report, released the same week, found that 79% of organizations now run AI agents in production while only 2% use identity security tools built to govern them: a 40-to-1 gap between deployment and identity discipline. Google just widened the top half of that ratio. Your move is the bottom half.
Before the agent coworker lands, settle five disciplines. They are the same five I have been arguing all along; the Gemini announcement just puts a deadline on them.
1. Identity. Provision the agent like you provision a person, because the directory will now treat it like one. A Workspace identity means an email that sends, a calendar that books, a Drive that owns files. That is a birth certificate with permissions attached. Write the onboarding: who authorizes creation, what role it is granted, what it can touch. Write the offboarding too. I have covered the full discipline in Every Agent Needs an Identity and Standing Privilege Is Dead: named identities, zero standing privilege, just-in-time elevation, revocation that actually revokes. The agent that can read everyone’s calendar needs an expiration date on that access, not a checkbox from install day.
2. Memory. The agent remembers. In the 4M reference architecture from my corpus work, memory comes in two tiers: session memory manages a single interaction (history windowing, running summaries), while persistent memory retains knowledge across sessions in a durable store. Google’s agent carries “the same memory everywhere,” which means the persistent tier is now a first-class enterprise asset. Treat it as one. Decide what the agent is allowed to retain, who can read that store, how long entries live, and what happens to the memory when the agent is reassigned. A coworker who changes teams hands back the sensitive files. Your agent should too. Memory that never expires is an undeclared archive of everything the agent ever saw, and archives get subpoenaed.
3. Permissions. Email, calendar, and Drive translate to four questions, the same ones from The Permission Is the Policy: what can it see, what can it send, what can it change, and what could it reconstruct by combining the other three? The last question is the one people skip. Read access to the calendar plus send access to email equals a meeting invite nobody scheduled, sent to the whole company, that looks like it came from the agent’s owner. Then add the MCP connection: Google’s agent can plug into any Model Context Protocol server. The Enterprise MCP Guide 2026, released October 5, catalogues six attack classes against MCP, including tool poisoning, tool shadowing, and shadow servers, and notes 68 MCP server CVEs landed in a single month. Every MCP server your agent connects to is a vendor with code execution inside your boundary. Run it through the permission review like a person, and the tool allowlist like an app inventory.
4. Spend. Google built the right primitives: multi-model orchestration, Smart Routing, and real-time spend caps. Caps you never set are decorations. Persistent coworkers run long-lived tasks and coordinate subagents, which is exactly the topology where costs compound. The practitioner angle is unchanged from Every Agent Needs a Budget: a per-agent budget enforced in the harness, token zones per task class, and a hard stop that fires before the invoice, not after it. The vendor gives you the dial. Turning the dial is still your job.
5. Accountability. The agent’s actions must land in an audit trail attributed to a named identity, under a named human owner. Agent Conduct Is Company Conduct covers what an investigator will ask for; the answer starts with “show me what the agent did, under whose authority, and who signed off.” A persistent coworker that learns from each interaction is also a coworker whose judgment drifts over time. That is a drift problem, and drift plus a mailbox is a liability pattern. The owner is the person who can say, six months from now, “this action was authorized, here is the record.”
None of this is an argument against the agent coworker. The agent that sits in the flow of work, with the context to act instead of waiting to be asked, is genuinely the shape of what comes next. But a coworker with an email address is an insider, and the enterprise already knows how to govern insiders: provision, bound, watch, and offboard. The discipline is written. The deadline is now.
Thinx-Tank