Google brings agentic AI to Gemini, starting with businesses

Google announced on Thursday (October 8), at a Google Cloud event, the arrival of a unified AI agent in Gemini: a system that not only answers questions but executes tasks on the user's behalf from a single interface. The news, reported by TechCrunch, arrives first for businesses that already use Gemini Enterprise, ahead of any rollout to the end consumer.

The choice is not accidental. According to Google CEO Sundar Pichai, Gemini already has more than 1 billion monthly active users, and about 90% of Fortune 100 companies use Gemini Enterprise at work. Pichai said starting with the corporate environment lets the company solve "the hardest problems of security, scale, and performance" that come with agents capable of acting on their own first.

From conversing to executing

The move follows a broader shift in the AI market: tools that have stopped being merely conversational and started taking on entire tasks, such as generating code, scheduling meetings, booking trips. TechCrunch cites as backdrop Meta's Muse, Instinct, and ChatGPT's newly launched Dots, all part of this same race for agents that operate within messengers and real workflows.

For software developers, the difference between an assistant that suggests and an agent that executes is operational, not cosmetic. Thomas Kurian, CEO of Google Cloud, summed up the change in approach:

Objectives, not just instructions.

Thomas Kurian, CEO of Google Cloud

In practice, according to Kurian, this means the agent plans the work, uses custom skills and tools, and connects to the company's internal systems to fulfill an end-to-end objective, rather than just following a single isolated command.

A team member with its own Workspace account

The most unusual detail of the announcement is how the agent is treated within the organization: it gets its own account in Google Workspace, as if it were just another employee. This includes its own email address and its own context: the agent knows who is on which team, time zones, who needs to approve what, and what's on people's calendars.

Users can trigger the agent by tagging it (@), sending it an email, sharing a document with it, or adding it to a chat group. Every action it takes generates an audit record attributed to the agent itself, not to a person, which solves (at least in theory) a recurring traceability problem in AI automation within companies.

Work progress is visible in an interface called the "tasks inbox," where you can follow the agent's reasoning, the delegation of tasks to subagents, the loading of specific skills, and the code generated in real time.

Where it connects: Git, BigQuery, Jira, and MCP

For those who work with data and infrastructure, the list of integrations is the most relevant point of the announcement. The agent connects to the company's internal tools, including:

• Google Workspace and Microsoft 365
• Slack, Jira, and Confluence
• Git
• BigQuery, Databricks, Postgres, and Snowflake

In addition, the agent works with any MCP (Model Context Protocol) server, inside or outside the company's network. This puts it in the same integration ecosystem already being adopted by other agentic tools, and it means teams that already expose data and skills via MCP for other assistants don't need to rebuild that layer from scratch for Gemini.

Requests to the agent can also include attachments, such as files, folders, or entire projects assembled for specific workflows, combining files and skills in a single request.

Model choice: Gemini by default, Claude as an option

By default, the AI chooses the best model on its own to complete each task. But users can take control and manually switch models, including third-party options, starting with Anthropic's Claude models. Google said it will expand this model selector to include open-source models and other private models in the future.

For those already operating on Vertex AI, this formalizes something many teams were already doing manually, orchestrating calls between Google's own models and third-party ones. The difference is that now this model choice happens inside the same orchestrating agent, rather than as separate integrations maintained by the engineering team.

Where the agent runs: from the terminal to Slack

The agent will be accessible on iOS and Android, Windows and Mac, Google Workspace, Microsoft 365, ServiceNow, Slack, and also from the command line. That last entry point is the one of most interest to those who write code: it means the same agent that delegates administrative tasks via Slack can also be invoked from the terminal, in the same workflow used to run tests or deploy.

Early testers included sportswear brand On, Shopify, and PayPal. Among the Gemini Enterprise customers cited by Google are BNP Paribas, Bradesco, Merck, Orange Spain, Santee Cooper, SOMPO, Ulta Beauty, and Wesfarmers. Bradesco's presence on the list is the clearest sign that the feature isn't just an experiment limited to mature markets: Brazilian banks are already part of the user base set to receive the agent.

Cost: orchestration, routing, and spend caps

Running multiple models for open-ended tasks has an obvious risk: unpredictable cost. Google said it will offer flexible spending options, including multi-model orchestration, intelligent routing between models, and real-time spend caps, to help companies keep the cost of running agentic AI in production under control.

For engineering teams that currently monitor token spend per project on Vertex AI, this real-time cap control is the most practical item in the package: it shifts part of the cost-governance work currently done with scripts and internal dashboards into the platform itself.

What remains unclear

Google did not detail, in TechCrunch's coverage, timelines for general availability, specific pricing for the agent, or how data policy works when the agent accesses sensitive systems such as production BigQuery or Postgres instances. There is also, so far, no announced date for the consumer rollout mentioned by Pichai. These are points that will determine whether the agent becomes a daily-use tool for Brazilian engineering teams or remains limited to controlled pilots among the large accounts that already use Gemini Enterprise.