Small teams, short runway, and pressure to ship faster are the context Google for Startups gives for the new Gemini agent, introduced at Gemini at Work 2026. The company also made its frontier model Gemini 4 Argon widely available, trained to maintain context across long, multi-step tasks. Instead of prompting separate chatbots for writing, coding, and data analysis, the Gemini agent is described as a single agent that takes high-level objectives and returns finished work. It is defined by three principles: unified, omnipresent access, and persistent. For complex jobs, it dynamically spins up temporary, job-specific sub-agents and coordinates their work in parallel or in sequence; it works across the web, command line, desktop, mobile, and natively inside Gmail, Docs, Sheets, Slides, Drive, Chat, and Calendar. It runs in the cloud with continuous memory across session, semantic, procedural, and episodic dimensions, so it can keep multi-day tasks moving after you close your laptop and onboard itself like a new hire.

Five startup takeaways are highlighted. First, coworker agents and skills can scale output: founders can create persistent coworker agents that have their own Workspace account, email address, Drive storage, and directory presence. These can be assigned roles such as launch coordinator, QA triager, or financial analyst, and teammates can delegate work by adding agents to a Chat space or @mentioning them in a Google Doc comment. Agents can be equipped with reusable skills, or Gemini can write skills for itself as it learns processes. For example, operational reporting skills connected to BigQuery and the Knowledge Catalog let anyone on the team ask a plain-language question to build and save a live report; saved reports run on demand without additional token costs. Pre-built specializations start with legal and financial services.

Second, open and third-party model choice means Gemini is the agent while the model underneath is the user’s choice. Workloads can be dynamically routed across Gemini models, Anthropic’s Claude, and lightweight open models like Gemma, letting startups optimize cost per task and swap models without rebuilding workflows. Third, the agent connects to existing services and tools beyond Workspace: Git, Slack, Jira, BigQuery, Postgres, Snowflake, Databricks, local desktop files, and any Model Context Protocol (MCP) server inside or outside the network for custom internal tools.

Fourth, built-in security and governance are aimed at B2B procurement: each agent gets a cryptographic Agent Identity for audit trails, runs in an isolated sandbox, and has built-in prompt injection defenses. External traffic routes through Agent Gateway so governance policies can be enforced, and coworker agents are managed much like users because they all have a Workspace account. Fifth, real-time cost controls let startups set hard spend caps directly in the Cloud Billing Console; if a cap is reached, the project’s agent automatically pauses so a runaway loop does not create a billing surprise. The article also notes Gemini runs on Google’s specialized AI infrastructure, including inference-optimized TPU 8i, which is said to deliver 80% better performance-per-dollar than the prior generation.

Startups are already using the platform. Whenere used Dreamlands, an agentic 3D worldbuilding platform built on the Gemini platform, to bring Pride and Prejudice to Fortnite. Gemini generated textures for each new object in about seven seconds for under ten cents, and one artist created nearly 250 custom assets in a single day and completed a historically accurate estate in under a week. Other named builders include Lovable, OpenEvidence, Recursion, and Replit.

The new Gemini agent capabilities are not publicly available yet. Google for Startups says startups can contact its sales team to learn more and discuss readiness, and can explore the Google for Startups program for credits, technical guidance, mentorship, and founder community programs.