Skip to content

gemini becomes an agent for work – the app has 1b users

A green robot in a suit and a man play chess with a stapler and a laptop in a graffiti basement office ai agents

google cloud's gemini at work keynote on oct 8 introduced the gemini agent, one agent for work that plans tasks, runs in the cloud and connects to company systems. google said the gemini app has one billion users, nearly 90% of the fortune 100 use gemini enterprise, and nearly 500 customers each processed more than one trillion tokens in the past year.

gemini at work '26 in 58 seconds: gemini becomes an agent, argon benchmarks, smart storage and gemma on a satellite.

what google announced

  • gemini agent – one agent for chat, knowledge work, images, media and code. it runs in the cloud, so a job keeps going after you close the laptop. the same agent works on web, mobile, desktop, the command line, google workspace, microsoft 365, servicenow and slack, with one shared memory.
  • roster of agents – on a larger job, gemini creates agents, gives each a task, messages them and coordinates the result.
  • coworker agents – in workspace an agent can have its own account and email address, and it sees only what a team shares with it.
  • skills and tools – skills are reusable instructions published to a shared registry. tools connect the agent to workspace, microsoft office, slack, jira, confluence, git, salesforce, servicenow and any mcp server.
  • data – a knowledge catalog defines business terms such as "net margin" once for all agents. smart storage uses gemini to enrich unstructured files in place. the borderless lakehouse queries amazon s3 and azure data lake with no variable egress fees.
  • industries – gemini for financial services and for legal is in preview. government, healthcare and retail are coming soon.

controls

  • identity – each agent has its own identity, cryptographically attested, with least-privilege permissions.
  • audit – every action is logged under the agent's name, not the user's.
  • budget – a limit is set before the agent starts and covers tokens, code sandboxes and compute. caps are tracked per project, and the agent pauses when the cap is reached.
  • network – agents run in a sandbox, and traffic between agents goes through agent gateway, an ai firewall that checks prompts and tool calls against one policy.

models and cost

the agent can run a job on argon, flash or the latest models from anthropic, and google says other proprietary and open models are coming. google's post puts it this way: "gemini is the agent, and the model underneath it is a separate choice."

thomas kurian said per-token prices have fallen 98% since 2024, while enterprise ai spending keeps rising because one complex agentic task can use 100,000 tokens. smart routing starts a job on a cheaper model and switches to a frontier model only when the job needs it. on stage he said google's own tests came close to running everything on a frontier model at a third of the cost and 40% faster. the written post does not repeat that figure.

the numbers

benchmarkgemini 4 argonnext best on the slide
vals index (knowledge work)68.9%67.0% (claude sonnet 5.5, claude opus 5.5)
deepswe v1.1 (coding)77.9%74.2% (claude opus 5.5)

we collected these figures from the benchmark slides google showed at the keynote. the slides mark the data as of oct 6, and we have not tested it ourselves. google also said argon and antigravity drove a more than 35% increase in agentic code submissions inside google over four weeks.

customer results from the keynote, as reported by google:

  • nokia – network troubleshooting time cut by up to 80%.
  • ntt docomo – time from data to insight cut from two weeks to instant, 450,000 hours a year returned.
  • snap – diagnostic troubleshooting cut from 30 minutes to 30 seconds with an agent on its storage archives.
  • lloyds banking group – close to 1,000 engineers on its gemini enterprise platform.
  • merck – about 75,000 employees with access to gemini enterprise.
  • nasa jpl – gemma, google's open model, runs on a satellite in orbit, a first for a vision language model.

in the finance demo, the agent took a video of a $48b all-stock buyout and a fund's live positions, flagged a breach of a 25% energy limit by $1.625m and recommended trades to fix it.

what google did not say

every benchmark, customer result and savings figure in the keynote comes from google or its customers, and the written post gives no methodology. it also gives no general availability date, plan tiers or pricing for the gemini agent. sundar pichai said the consumer version is planned "soon."

sources

ON AIR · RADIO.THEHYPE.NEWS ↗ ai news radio — 24/7