Give the work agent its own identity and a spend cap

On October 8, 2026, Google Cloud announced the Gemini agent. Google says you hand it an objective and that it plans the work, uses skills and tools, connects to your systems, and returns finished output in the documents, the inbox, and the developer tools. Google says people can open it on the web, a phone, a desktop, the command line, Google Workspace, Microsoft 365, or Slack, with or without a screen of its own, and that work lasting hours or days keeps running after the laptop closes.

A center of excellence can start with three controls Google attaches to that agent: its own identity, an audit trail under that identity, and a project spend cap that pauses the work. The model underneath is a separate choice. Google says the job can run today on a Gemini model or on Claude from Anthropic, that the skills, the context, and the data stay put when the model changes, and that matching the model to the job lowers cost on the simple jobs.

Whose name is on the work

Google says a sub-agent is temporary, tied to one job, and has its own identity, and that a coworker agent keeps a standing role across days. Google says a coworker gets its own email in the form @agents.company.com, its own storage, and only the context a person or a team provides.

Inside Google Workspace, Google says you describe the role and the product creates an agent with its own account: email, calendar, Drive, and a listing in the company directory. Google says colleagues add it to a chat or mention it. Google says that, tagged in a document comment, it can suggest an edit and appear under its own name in version history, that it sees only what was shared with it, and that access follows the sharing the team already uses.

Google says departments can publish skills, reusable multi-step instructions, to a company registry, and teams can publish tools to a second registry. Google says the agent can connect to a Model Context Protocol server. Google says the agent chooses the skills and tools.

The log and the cap

Google describes four controls. Google says that identity is cryptographically attested, that the agent is governed like an employee, and that security administrators approve its permissions. Google says the identity is written into the logs, into any virtual machine the agent starts to run code, and into outside connections through standards such as OAuth.

Google says every action goes to an audit trail attributed to the agent rather than to a person, that agents run in a sandbox with its own network boundary, and that traffic in, out, and between agents passes through a gateway. Google says that policy is written once and applied to every agent. Google’s example is one sentence: “agents may not open documents classified Need to Know.”

Spend sits beside those controls. Google says you can set a hard limit on a project’s AI spend in the Cloud Billing Console. Google says it watches token use and sandbox cost, pauses that project’s agent when the cap is hit, and lets a person resume it from the console. The cap is per project, and Google says a company can charge the spend back to a department.

Two adoption figures on the post are company-reported. Nearly 80 percent of Google Cloud customers use Google’s AI products, and nearly 90 percent of the Fortune 100 use Gemini Enterprise. Neither figure says whether any one process has an identity, a log, and a cap.

Six lines for one process

Fill this card before connecting an agent to a live system. The same day, Orange Spain named incident triage and published no permissions, log, or cap for it. Leave those lines blank until your own console shows them.

1. Identity. The agent’s directory name or email. For a Workspace coworker, include calendar, Drive, and the directory listing.
2. Allowed actions. The steps this process may take. Mark the step a person still performs, such as sending a message or changing a record.
3. Audit. Where the action appears under the agent’s identity, and who can open the log.
4. Never touch. One line in the form of Google’s example. Name the mailboxes, systems, or document classes closed to this agent.
5. Spend cap. The Cloud Billing project, the limit, and who may resume the agent after a pause.
6. Model. The model for this job, on its own line.

Open the console and look for line 3 and line 5. The test passes when both are visible. Until they are, the send step stays with a person. The card records what the console shows. Safety, completeness, and fitness for a regulated decision stay separate questions.

The same day’s count

Orange Spain and Google Cloud said Orange has deployed more than 1,000 custom agents on Gemini Enterprise. Employees built them on a low-code framework across human resources, IT, sales, and customer service. The companies report a near 100 percent Gemini Enterprise license activity rate within 30 days of assigning licenses across major business areas. Both figures are company-reported. Named uses include incident triage, root-cause analysis of system logs, finding internal knowledge, and routine internal requests.

The companies describe employees building the agents, with centralized control of orchestration. A skill from a non-technical team still needs the six lines.

What to do this week

• Pick one process a team already wants an agent to run.
• Fill the six lines where the team can see them.
• In Cloud Billing, confirm that project’s spend cap. Google says a hit pauses the agent.
• For a coworker agent, write who may add it to a chat.
• Name who approves a new skill on the company registry.

Filling the card for one live process is the work of a consult block: which process moves, whose name the agent uses, and what dollar figure pauses it. Consult is $400 an hour, in five-hour blocks.