Google pitches Gemini agent to startups for busywork

Google has outlined how startups are using its Gemini agent for product, operations and reporting tasks. The system is designed as a single agent that works across Google Workspace and external tools.

The announcement comes as Google also pushes broader adoption of Gemini 4 Argon, which it describes as a model built to hold context across long, multi-step tasks. Together, the products target founders and small teams looking to spend less time on administration, coding support and internal analysis.

Google says the Gemini agent is meant to replace separate chatbots for writing, coding and data analysis with one system that can take a high-level instruction and return completed work. It can handle knowledge work, generate creative assets, analyse data, write and run code, and create temporary sub-agents for more complex jobs.

The agent is available on the web, command line, desktop and mobile, and is built into Gmail, Docs, Sheets, Slides, Drive, Chat and Calendar. Google says it retains memory across sessions and can continue working on tasks over several days.

Team roles

A central part of the startup pitch is the use of so-called coworker agents. Businesses can create persistent agents with their own Workspace account, email address, Drive storage and directory presence, then assign them roles such as Launch Coordinator, QA Triager or Financial Analyst.

Teams can add those agents to Chat spaces or mention them in document comments to take on specific tasks. They can also equip them with reusable skills or let Gemini create those skills as it learns internal processes.

One example is operational reporting linked to BigQuery and the Knowledge Catalog, allowing staff to ask a plain-language question to create and save a live report. Google says saved reports can then run on demand without additional token costs.

Model choice

Google says the Gemini agent is separate from the underlying model, allowing customers to route tasks between Gemini models, third-party systems such as Anthropic's Claude and open models including Gemma. The approach reflects a wider shift among startups towards mixing models to manage computing costs and avoid relying on a single supplier.

Google argues that sending every task to a frontier model can hurt margins as a business scales. The agent lets teams reserve more advanced models for complex reasoning while using smaller models for repetitive work.

Tool access

Google also positions connectivity as a key part of the offering. The agent can connect to development and collaboration services such as Git, Slack and Jira, as well as data sources including BigQuery, Postgres, Snowflake and Databricks.

It can also connect to local desktop files and custom tools through the Model Context Protocol. Google says that means engineering teams can link internal tools to the agent without maintaining bespoke connectors for each workflow.

Security focus

For business customers, Google highlights governance and audit controls. Each agent is assigned a cryptographic Agent Identity, runs in an isolated sandbox and includes defences against prompt injection. External traffic is routed through what Google calls Agent Gateway, which it says allows governance policies to be enforced.

The focus is aimed at business software buyers that subject AI systems to procurement and security reviews before signing contracts. Google says coworker agents can be managed in the same way as other Workspace users.

Cost controls

Google says customers can set real-time spending caps in the Cloud Billing Console so an agent pauses automatically if a project reaches its limit. The measure is meant to prevent long-running autonomous tasks from generating unexpected charges.

Under the service, Gemini runs on Google's own AI infrastructure, including the TPU 8i. Google says the chip offers 80% better performance per dollar than the previous generation.

Startup use

Google pointed to early startup use cases including Dreamlands, a 3D worldbuilding platform built on Gemini. It said the company used the system to create textures for new objects in around seven seconds for less than 10 cents each, enabling one artist to create nearly 250 custom assets in a day and complete a historically accurate estate in under a week.

Google also named Lovable, OpenEvidence, Recursion and Replit among businesses working with its AI tools. The latest Gemini agent features are not yet publicly available, but startups can discuss access and preparation with Google teams.