Google's work agent gives every team a coworker with an inbox

Google Cloud launched a single agent for work on October 8, and the headline feature is not the agent. It is the coworkers.

The Gemini agent is Google's pitch for where enterprise agents actually live: one system that plans work, uses tools, connects to business systems, and returns finished results inside documents, inboxes, and developer environments. It works inside Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, and it can also be reached through Microsoft 365 and Slack. That much is a conventional launch. The unusual part is what Google lets you create inside it: "coworker agents" that act as team members, with their own email address and access only to the information they are given. For multi-step work it can create temporary sub-agents with distinct identities and coordinate them in parallel or sequence; Google says it maintains memory and context across devices and channels and can continue work for hours or days after a user closes a laptop. Per the company, coworker agents receive only the context shared with them. Google's own materials mention activity logging alongside agent identity and role-based permissions; those controls, per RuntimeWire's read, "will determine whether companies let software act across sensitive systems rather than simply draft a response." One caution that outlet adds and that should be said plainly: the announcement "supplies no independently tested comparison of quality or savings."

This lands in a month that is rapidly filling with competing answers. OpenAI's Dots are the always-on agents that chase goals across apps. Meta's Muse is the personal agent that shops, books, and pays. Anthropic just pushed Claude into Google Workspace as a first-party beta, and now Google is answering with the whole enterprise stack: objectives, skills, tools, and a model picker. The picker is the detail to watch. By default the agent selects the best model for each task; users can override, starting with Anthropic's Claude models, with open-source and other private models to come. Google is selling the agent layer above the model war, and letting you run a rival's model inside it.

The enterprise pitch is sequenced deliberately. CEO Sundar Pichai framed the launch around Google's distribution: over 1 billion monthly Gemini users, nearly 90% of the Fortune 100 on Gemini Enterprise, per the company's announcement. Thomas Kurian, the Google Cloud CEO, said the business-first rollout exists to solve "the harder problems around security, scale, and performance" before any consumer release. The governance story the company tells: identity and policy management, authorization and permission controls, secure sandboxing and network gateways, plus cost controls via multi-model orchestration, Smart Routing, and real-time spend caps. The agent can connect to company software and to any MCP server, use custom skills, and, per the post, learns from each interaction. Google's stated tool reach, per RuntimeWire: Microsoft Office, Teams, Slack, Confluence, Git, Jira, Salesforce, ServiceNow, plus BigQuery, Databricks, Postgres, and Snowflake, with CLI and third-party-app operation. Industry versions for financial services and legal are in preview; government, healthcare, and retail follow.

Stripped of the launch language, Google is making three bets.

Bet one: the agent layer is multi-model. This is the most interesting commercial claim in the announcement. If the model picker is real and neutral, Google becomes the Switzerland of workplace agents, and every model lab's differentiation narrows to the picker. That is exactly what an operator should verify first. "Picks the best model for each task" is a claim about routing quality, cost accounting, and audit trails. Who decides which model ran on a given task? Does the audit log say so? Do spend caps attach per model or per agent? The launch materials say cost controls exist; they do not say what they cost to run or what happens when a cap trips mid-task. Multi-model orchestration is a governance feature only if the orchestration decisions are inspectable.

Bet two: identity is the unit of agent provisioning. The coworker agent with its own email address is a bigger idea than it looks. An agent with an email address can be invited to meetings, added to groups, CC'd on threads, and, crucially, offboarded. That maps onto identity infrastructure operators already have: provisioning, deprovisioning, access reviews, scoped entitlements. It is also where the failure modes live. "Access only to the information they are given" is a promise about every system the agent touches, including the ones connected through an MCP server or a custom skill. The first question for any pilot: when you revoke a coworker agent's access, does anything it already read stay read, and who sees what it wrote? The second: what does its memory retain across tasks, and is that memory per agent or per user? The launch materials do not answer these yet.

Bet three: long-running work is the job, not the demo. The agent works after you close the laptop, coordinates sub-agents, and returns work inside the applications people already use. That is the difference between an assistant and a worker, and it is where the audit-log problem from the OpenAI rogue-agent reviews comes back. If OpenAI needed months and 50 petabytes to reconstruct what its own agents did, the first requirement for any long-running enterprise agent is that someone can reconstruct it. Google's pitch includes the controls; the pilot question is whether the reconstruction is built in or bolted on.

What remains unknown: pricing, which Workspace tiers get the agent and when, whether coworker agents are GA or preview, what Smart Routing actually optimizes for, and whether the third-party picker extends beyond Claude in practice. The named early testers (Shopify, PayPal, Bradesco, Orange Spain) are company-reported anecdotes, not independent evaluations. Treat them as leads for who to ask, not as evidence.

The operator move: this deserves a pilot, not a procurement. Scope it to one team, one set of scoped coworker agents, with an explicit kill switch on memory retention and a standing export of the audit trail. Ask for the model-routing log before the demo. The platform pitch is genuinely strong. The governance questions are the whole story, and they are not answered in a keynote.