AgentsAutonomy & tool use
Google launches Gemini as a work agent for businesses, with its own Workspace account and audit trail
Google's Gemini agent handles chat, multi-step tasks and code from one interface, delegates to subagents, runs on Gemini or Claude models, and gets its own Workspace identity and audit trail. It is in private preview for businesses, with consumers later.
Reach: not measured yet
AT Pulse research deskPublished · 5 min read · AI-assisted analysisGemini is being turned into one agent for chat, tasks and code. It works on desktop, web, CLI, mobile and browser, and inside Workspace, Microsoft 365, ServiceNow and Slack. Because it runs in Google's cloud, every surface shares one memory, and larger jobs can be split across subagents it creates as needed. It is in private preview for businesses, with consumer access planned later. [1] [3] [4]
Each agent has its own identity, separate from the user's, including a Workspace account and email address. Its actions are logged under its own name. Admins can give it least-privilege permissions and set project spend caps that cover tokens, sandboxes and compute. Tasks run in an Agent Sandbox, and all traffic passes through an Agent Gateway that enforces policy. [1] [3]
The agent and its model are separate choices. Gemini 4 Argon, Flash and the latest Claude models are available now, and other models are promised. Google's own testing says Smart Routing, which starts on cheaper models and moves up only when needed, gets close to running everything on a frontier model at a third of the cost and 40% faster. [3] [5]
each agent gets separate identity, policies, audit logs and project spend caps.
Read the full assessment
Implication: enterprises could govern agents through existing access-management and chargeback processes, while model choice reduces dependence on one provider, though routing quality needs workload-specific testing.
Executive brief
Google is giving its new Gemini work agent a Workspace account of its own: an email address, a calendar and an audit trail that records actions under the agent's name, not the user's. The agent was announced at Google Cloud's Gemini at Work event on October 8, 2026. It takes in chat, multi-step tasks and code from one interface, can hand work to subagents, and can run on Gemini or Anthropic's Claude models. Google is launching it to businesses first, in private preview, and says consumers will get it later (TechCrunch; 9to5Google). Most of the performance and customer figures come from Google.
What changed and event timeline
Gemini 4 Argon announced
Google's new frontier model went first to cyber defenders in its Fairwind Program. It is priced at $2/$10 per million input/output tokens during an introductory period (). The Hacker News gives the release date as October 1 ().
- Early Oct 2026
Smart Routing and spend caps
Thomas Kurian, CEO of Google Cloud, said these cost controls were introduced "earlier this month." No exact date was given ().
Universal Gemini agent unveiled
Sundar Pichai, CEO of Google and Alphabet, presented Gemini as a "universal agent for work" and said it is launching to businesses before consumers (;).
Private preview; industry versions
The agent is in private preview, with wider availability promised "soon" for some Workspace plans (). Versions for financial services and legal are in preview. Government, healthcare and retail versions are listed as coming soon ().
Capabilities and access
- One agent for three kinds of work. It handles chat, tasks and code across desktop, web, CLI, mobile and browser.
- Gemini 4 Argon, Flash and the latest Claude models are available now. Google says other proprietary and open models are coming (keynote).
- These include Jira, Confluence, Git, BigQuery, Databricks, Postgres and Snowflake, plus any MCP server (TechCrunch).
Read the full section
- One agent for three kinds of work. It handles chat, tasks and code across desktop, web, CLI, mobile and browser. It also runs inside Workspace, Microsoft 365, ServiceNow and Slack (keynote).
- Models. Gemini 4 Argon, Flash and the latest Claude models are available now. Google says other proprietary and open models are coming (keynote).
- Connectors. These include Jira, Confluence, Git, BigQuery, Databricks, Postgres and Snowflake, plus any MCP server (TechCrunch).
- API. An API lets other products use the agent without its own interface (9to5Google).
- Pricing. Google has not announced pricing. It says it does not plan per-seat fees for coworker agents (Futurum).
Technical analysis for researchers and developers
- Runs in the cloud. Because the agent runs in Google's cloud, every surface shares one memory and one "personalization graph."
- The agent keeps four kinds of memory: session, semantic, procedural and episodic (keynote). A "Dream Mode" turns past conversations into stored preferences (keynote).
- Skills are reusable instructions stored as markdown files or modular prompts and published to a shared registry.
Read the full section
- Runs in the cloud. Because the agent runs in Google's cloud, every surface shares one memory and one "personalization graph." Large jobs are split across a roster of subagents that it creates as needed (keynote).
- Memory. The agent keeps four kinds of memory: session, semantic, procedural and episodic (keynote). A "Dream Mode" turns past conversations into stored preferences (keynote).
- Skills. Skills are reusable instructions stored as markdown files or modular prompts and published to a shared registry. That is close to existing agent-skill conventions (keynote).
- Execution and governance. Tasks run in an Agent Sandbox. Traffic passes through Agent Gateway, which Google calls an AI network firewall that screens for malicious prompts and tool calls. Each agent has a cryptographically attested identity (keynote).
- Evaluation. No formal evaluation methodology for the agent itself has been published.
Claims and evidence
- Adoption: Over 1 billion Gemini monthly users, and nearly 90% of the Fortune 100 using Gemini Enterprise (TechCrunch; keynote). These are Google's figures.
- Argon benchmarks: 77.9% on DeepSWE v1.1 and 68% on CWE-bench v1 (Google blog). Google ran these; the reviewed sources include no independent replication.
- Internal coding gains: A more than 35% rise in agentic code submissions, and code rollbacks down by nearly a third (keynote).
Read the full section
- Adoption: Over 1 billion Gemini monthly users, and nearly 90% of the Fortune 100 using Gemini Enterprise (TechCrunch; keynote). These are Google's figures.
- Argon benchmarks: 77.9% on DeepSWE v1.1 and 68% on CWE-bench v1 (Google blog). Google ran these; the reviewed sources include no independent replication.
- Internal coding gains: A more than 35% rise in agentic code submissions, and code rollbacks down by nearly a third (keynote). These are internal Google figures.
- Smart Routing: Results "close to" running everything on a frontier model, at a third of the cost and 40% faster, "in our testing" (keynote).
- Customer results: PayPal routes 10 million multi-model requests a week (keynote). These are customer anecdotes from the stage.
Context and prior work
- Rival agents: TechCrunch places the launch alongside Meta's Muse and ChatGPT's Dots (TechCrunch).
- Microsoft: Futurum names Microsoft Copilot, already built into Microsoft 365, as the main competitive obstacle (Futurum).
- Gemini Enterprise: The agent builds on Google's existing Gemini Enterprise platform (Wikipedia).
Read the full section
- Rival agents: TechCrunch places the launch alongside Meta's Muse and ChatGPT's Dots (TechCrunch).
- Microsoft: Futurum names Microsoft Copilot, already built into Microsoft 365, as the main competitive obstacle (Futurum).
- Gemini Enterprise: The agent builds on Google's existing Gemini Enterprise platform (Wikipedia).
- Gated model release: Argon's cyber-first rollout fits the industry pattern of releasing frontier models to selected users first (The Hacker News).
Limitations, safety and contested findings
- Argon access is restricted. Its broader rollout depends on stronger safeguards. Google says it watches Argon's chain-of-thought and actions and stops execution when needed.
- Timelines and scale are unproven.
- The on-stage ML demo reported 90% accuracy and 85% precision on historical data, with no details about the dataset (keynote).
Read the full section
- Argon access is restricted. Its broader rollout depends on stronger safeguards. Google says it watches Argon's chain-of-thought and actions and stops execution when needed. It also plans a version without cyber safeguards for trusted defenders (The Hacker News).
- Timelines and scale are unproven. Futurum notes that few general-availability dates were given and that grounding the agent in company data is unproven at enterprise scale (Futurum).
- Demo metrics lack context. The on-stage ML demo reported 90% accuracy and 85% precision on historical data, with no details about the dataset (keynote).
- Release date differs. Argon's date is September 30 on Google's blog and October 1 at The Hacker News.
Business and practitioner implications
- Treat agents as staff for governance.
- Spend caps cover tokens, sandboxes and compute, which makes chargeback to departments possible (keynote).
- The model picker reduces dependence on one provider, but routing quality needs testing on each organization's own workloads.
Read the full section
- Treat agents as staff for governance. Each agent has its own identity, least-privilege permissions and audit log, so existing identity and access management (IAM) processes can cover it (keynote).
- Budget per project. Spend caps cover tokens, sandboxes and compute, which makes chargeback to departments possible (keynote).
- Less model lock-in. The model picker reduces dependence on one provider, but routing quality needs testing on each organization's own workloads.
- Early adopters. On, Shopify and PayPal tested the agent early. Google says over 50 companies run it today (keynote).
The source trail.
Sources (8)
Google brings agentic AI to Gemini, starting with businesses
Article text retrieved; extracted text may omit tables or interactive elements. A transcript of the video linked from this page is supplied (linked video manual captions; https://www.youtube.com/watch?v=1ASzWklab2U); automatic text may contain errors.
techcrunch.comGoogle is launching a one-stop Gemini agent for your work tasks
Related coverage; assess separately
theverge.comLinked video: Google brings agentic AI to Gemini, starting with businesses
linked video manual captions
www.youtube.com