Business & future work · Inquory Research

Google announces a work agent with its own identity: check who can act

Google announced a Gemini work agent with a distinct identity. Separate acting authority, delegated access, activity logs and tested enforcement before handing over a workflow.

· Sources checked October 9 · AI-assisted research and drafting · Editorially reviewed · All five sources are Google publications · No account, rollout, enforcement, customer outcome or savings test

A task token with a blank identity badge carries an envelope toward a closed gate while a separate paper roll records activity and a reviewer holds an unfinished checklist.
Conceptual illustration of a delegated task, an activity record and a permission boundary. A recorded action does not by itself demonstrate that a restriction was enforced.

The blank badge, separate activity record and closed gate distinguish identity, observation and enforced refusal without depicting Google software or a tested customer system.

AI-generated editorial illustration · Inquory

By Inquory Research

Published October 9, 2026. Sources checked October 9. AI-assisted research and drafting; editorially reviewed. Inquory has not tested the announced agent, a customer account or its enforcement controls.

Google announced a universal Gemini agent for enterprise work on October 8, describing a service that can carry out delegated tasks in the cloud and act as a Workspace coworker under its own identity. For a team deciding whether to hand over a workflow, the immediate distinction is between assigning an outcome and establishing the authority used to perform it.

What Google announced October 8

The Google Cloud post, adapted from CEO Thomas Kurian's Gemini at Work keynote, presents one agent for questions, knowledge work, media creation and coding. Google says tasks can continue after a laptop closes. Its Workspace account describes personal assistance, suggested delegation and a coworker agent with its own account; for that coworker, Google says access follows what the team shares. These are vendor descriptions, not observed behaviour in an Inquory test. October 8 announcement

The announcement also describes attributed agent activity, role-based authorization, sandboxing and a network gateway. It says project spending caps can pause an agent. We have not verified account rollout, Canadian availability, contractual scope or those controls in operation.

Inquory interpretation: a persistent worker creates an accountability question that a chat answer does not settle. Someone must identify the acting principal, approve the intended actions and review the resulting changes, including work completed while that person is away. A recognizable agent name in a conversation is a starting point for that record.

An agent identity and delegated user access are separate records

Google's Agent Platform documentation describes Agent Identity as supporting both an agent acting as itself and an agent acting on behalf of a user. Its authentication manager brokers outbound credentials, supports OAuth delegation and records access attribution. The guide also describes revocation of user-delegated permissions. This is documentation for the platform's identity mechanism, not confirmation that every announced Workspace surface exposes the same configuration. Agent Identity overview

For a team, the useful question is which authority a particular action uses. A file edit under an agent's own account, a tool request made with a user's delegated access and an internal proposal awaiting human acceptance should have separate entries in the team's record. Access to one source does not by itself establish approval to send its contents elsewhere. This distinction is our analysis, not an additional Google product promise.

A log does not demonstrate that a restriction was enforced

Google's Agent Gateway setup guide explicitly distinguishes two configurations. Audit-only permits traffic and generates logs without IAP enforcing IAM policies in that dry-run mode. Enforce policies blocks requests without an explicit Allow policy and is recommended there for production. The guide carries an October 8 UTC update. Gateway setup documentation

That is a distinction within an administrator-configured platform gateway. It does not establish a contradiction with the October 8 managed-agent announcement or prove the configuration of any company's Workspace agent.

Inquory interpretation: evidence of a recorded action answers an observation question. Evidence that an attempted action was refused answers an enforcement question. An approved evaluation should record both an expected allowed action and an expected refused action using fictional, non-sensitive material. These are proposed acceptance criteria; Inquory has not performed either test. This article provides no permission-change or deployment instructions.

Product eligibility and spending controls need their own scope

The developer-tool guide, updated October 8 UTC, has separate edition, invoiced-billing and regional conditions. It also says the Gemini Enterprise Admin role alone does not grant use of AI developer tools. Those requirements concern that feature; they are not a universal eligibility list for everything announced October 8. Developer-tool scope

The October 8 post links to billing background published August 26. That older article still qualifies pay-as-you-go app access as available to selected customers with broader rollout ahead, and deferred execution pricing as coming soon. It describes project caps that pause agent API calls, with separate resume and overage choices. Treat that page as background and obtain the terms applicable to the actual project before relying on a spending limit. Billing background

Fictional example: a draft estimate stays an internal task

Amina coordinates a fictional Canadian repair business. She wants an assistant to turn an internal job note into a draft estimate. Her proposed acceptance record identifies the approved test folder, the identity expected to produce the draft and the person who reviews it. It also specifies that the evaluation must not contact a customer or change a live record.

She separates three questions: did the draft reach the reviewer, did the record attribute its actions to the expected identity, and did the designated restriction refuse an out-of-scope request? A plausible draft leaves the latter checks unfinished. This is a fictional decision aid, not a customer result, available Canadian rollout or tested configuration.

For a small service business, this helps keep responsibility attached to the handoff: drafting, approving and communicating are different actions. No time savings, failure rate or traffic improvement is measured here.

What remains unresolved

All five sources are Google publications. They support an announcement and documented component behaviour; they do not independently establish reliability, an enabled account or successful enforcement. We have not established a complete edition-by-edition or region-by-region rollout schedule for the newly announced agent.

Google's enterprise-agent description should not be used to rewrite the separate personal-account, Workspace skills or Studio migration notices covered in our earlier report. The product and surface must be identified before treating a notice as applicable. A document's update date does not establish feature activation.

Related reading

Planned image credit: AI-generated editorial illustration — Inquory. The image is a conceptual depiction of identity, logging and a permission boundary, not Google software or a tested system.