An AI agent operates in an isolated environment with limited permissions and human oversight.

Gemini agent for business: the controls you should require from any AI agent

On October 8, 2026, at its Gemini at Work event, Google Cloud introduced the Gemini agent, which it describes as “a single, universal agent for work” (Google Cloud, 2026). Beyond the product itself, the announcement is useful for another reason: it details the governance controls a major vendor considers necessary for an agent inside a company. That list works for evaluating any agent, whether it comes from Google or not.

What Google announced

According to the article by Thomas Kurian, CEO of Google Cloud, you give the agent “objectives, not instructions.” Long-running tasks keep running in the cloud even after you close your computer. For multi-step jobs, the agent creates temporary sub-agents, each with its own identity (Google Cloud, 2026).

It connects to tools such as Slack, Teams, Microsoft Office, Salesforce, ServiceNow, Snowflake, BigQuery and Jira, and to any MCP server (the open protocol for connecting agents to systems). Users can access it from Google Workspace, Microsoft 365, Slack or the command line. The agent also orchestrates several models: currently those in the Gemini family and Anthropic’s Claude models.

In its announcement, Google did not publish a general availability date or pricing for the agent (Google, 2026). The article includes customer results, such as Bradesco, which reportedly cut document review from one hour to five minutes. These are Google’s claims, not independent measurements.

The governance controls

ControlWhat Google offers, according to its announcementWhat you should require from any agent
IdentityEach agent has its own identity, cryptographically attestedA dedicated account for each agent, never an employee’s
Least privilegeRole-based permissions, approved by security administratorsOnly the access the task needs, reviewed by someone accountable
LoggingEvery action is recorded in an audit trail under the agent’s nameA complete, searchable log, plus knowing who delegated the task
IsolationExecution in an isolated environment (Agent Sandbox) with its own network boundaryA space kept separate from critical systems
Traffic controlAll traffic passes through Agent Gateway, a firewall that enforces your policiesRules on which systems and data the agent can touch
Spending limitA per-project spend cap in the billing console; when it is reached, the agent pausesA spending cap set before you start, with alerts

Two important caveats. First, the spend cap is not instantaneous. According to Google’s documentation, it is applied to estimated costs, with some delay, and overages are billed as usual (Google Cloud, 2026). It became generally available on October 13, 2026, for certain services only (Google Cloud, 2026).

Second, the log attributes each action to the agent rather than to a person. That makes clear what the agent did, but not who asked it to. Your company must be able to answer that second question. Nor does the announcement describe a human approval step for specific actions.

What it means for companies in Ecuador

Many companies test agents using an employee’s account, with access to everything and no spending limit. That is convenient at first and risky in production. If an agent makes a mistake, you won’t know what it did or how much it cost.

Google’s announcement confirms what governance frameworks such as ISO/IEC 42001 or the NIST AI RMF already call for: identity, least privilege and traceability. We explain this in our guide to AI governance standards. And the problem is not theoretical: companies are adopting agents faster than they can oversee them.

What to do: a checklist for your agent

  1. Give it its own identity. Create a service account for each agent and each sub-agent.
  2. Limit its permissions. Start with read-only access and expand it once the agent proves it works.
  3. Log every action. Record what it did, with which data and who delegated the task.
  4. Isolate it. Run it in a separate environment and define which systems it can connect to.
  5. Set a spending cap. Configure a budget, alerts and an automatic pause; remember there may be some lag.
  6. Require human approval. Payments, messages to customers and changes to systems must go through a person.
  7. Vet the vendor. Ask about each of these controls before signing, whether it’s Google or another vendor.

At Wasivia, we design AI agents and multi-agent systems with their own identity, minimal permissions, a log of every action and human oversight. Read also AI agents: companies are adopting them faster than they can oversee them.

References

Similar Posts