Gemini Agent Brings Coding and Sub-Agents to Google Workspace

Google turns its assistant into an agent

Google Cloud used its Gemini at Work 2026 event on Thursday, October 8, to introduce the Gemini agent. The company calls it a single, universal agent for work. It can answer questions, handle knowledge work, create content and write code.915 Google Cloud CEO Thomas Kurian presented it in his keynote.2 CIO Dive reported that he summed up the change on stage by saying that Gemini now becomes an agent.11

The main pitch is delegation. Google tells users to give the agent objectives rather than instructions. The agent then plans the work, uses tools, connects to company systems and delivers finished output in documents, email and developer environments.89 Google's own example is a manager asking an employee for a project update as a slide deck. Workspace detects that this request can be handed off and offers a one-click option to pass it to Gemini.15

The coverage mostly agrees on the core facts, because most outlets worked from Google's blog post and Reuters' wire copy. Where reports differ, it is mainly in emphasis, plus a few details that have not been confirmed.

What the agent actually does

For anyone who follows AI productivity tools, the notable thing is how many product categories Google is folding into one agent. Writing assistants, meeting schedulers, research copilots and coding helpers have mostly been sold as separate products. Google is presenting them as skills of a single agent reached through one prompt box.8

The agent works inside Gmail, Docs, Sheets, Slides, Chat and Calendar. Users call it up by typing @Gemini.515 It can also be reached through Slack, Microsoft 365, the command line and an API. Developers can embed it in other software as a "headless" agent with no interface of its own.58 Google lists connectors to Jira, Confluence, Git, BigQuery, Databricks, Postgres and Snowflake.10 One outlet adds Salesforce, ServiceNow, Microsoft Teams and support for Model Context Protocol servers.16

The meeting-assistant use is the clearest early example. In Google's demo, a user asks Gemini to set up next week's meeting with "the usual" group of regional event leads and gives no names. The agent works out who they are from a chat space and an earlier email thread, checks their calendars and starts an email thread to agree a time. That includes people outside the company.515 Most standalone scheduling and meeting tools need the attendee list typed in. This demo relies on the agent already holding organizational context, which is something an add-on tool cannot easily copy.

Memory as the real product

Memory gets the most attention in coverage aimed at note-taking and research use. The agent runs in the cloud, so it keeps one set of memories and one personalization graph across devices and channels. A job that takes hours or days keeps going after the user closes the laptop.8 Google describes four layers of memory:515

• Session memory: the current task, even when it runs for days.
• Semantic memory: a structured knowledge base built from documents, conversations and other agents.
• Procedural memory: how jobs get done, including skills the agent writes for itself.
• Episodic memory: a record of everything the agent has done before.

Kurian described the benefit as never having to re-brief the agent or remember which machine a task was given on.1 In practice, this design puts the agent in competition with dedicated AI note-taking and knowledge-base tools. If the agent builds its own semantic memory from meetings, threads and documents, the case for a separate note-taking layer gets weaker inside companies already using Workspace.

XenoSpectrum argues that much of this is not new. Its timeline of Google announcements shows long-term memory and long-running execution arriving in April, continuous runs of up to seven days in July, spending caps in August and cross-app actions in Workspace in September.17 On that reading, October's news is less a technical breakthrough than a packaging decision: existing platform features are bundled into a work "colleague" with an ongoing role.17 That reading looks right to us. The launch is about putting the agent in front of ordinary employees, not about capabilities that did not exist before.

Coworker agents and sub-agents

Google describes two kinds of helper agents. For complex jobs, Gemini creates temporary, task-specific sub-agents, each with its own identity. They work in parallel or in sequence and report back.28 Separately, companies can set up persistent "coworker agents" for roles such as a team's project manager or an analyst in the finance department.315

Reports differ on how far that identity goes. Reuters' wire copy says only that coworker agents get their own email address and see only the information they are given.9 Other outlets describe much more: a full Workspace account with Calendar and Drive storage, an entry in the company directory, and edits recorded under the agent's name.1117 BigGo reports addresses in the format @agents.company.com, and says every action goes into an audit trail credited to the agent and that outbound traffic passes through an agent gateway described as an "AI network firewall."16 Taken together, the detailed accounts suggest Google intends agents to be first-class users in the directory. That is a bigger governance change than the short wire summaries suggest.

Gemini in name, multi-model underneath

The most surprising detail is that the Gemini agent does not run only on Gemini. Google says it picks the model that fits each job. Today that means its own Gemini models or Anthropic's Claude, with other private and open models to come.69 Google argues this improves quality and lowers costs. Real-time spending caps and smart routing are presented as enterprise cost controls.510

Inside AI reads the Claude integration as Google admitting that no single model is best at everything.4 XenoSpectrum stresses a practical benefit: business context, skills and system connections stay in place when the underlying model changes. It also notes that Google has not said how it chooses between models or how much routing actually saves.17 A 9to5Google reader raised a more basic point. If the agent is not tied to Gemini models, the branding risks confusing people about what Gemini is.15 There is a separate naming clash too: Google already shipped a consumer feature called Gemini Agent to AI Ultra subscribers in November 2025.19

Strategically, the move makes sense. Microsoft 365 and Slack support signal that Google cares more about interoperability than locking customers in, which suits large companies that use several vendors.4 In practice, Google is trying to win the coordination layer even when another company's model does the reasoning.

The competitive field

The timing places Google in a crowded field. OpenAI launched always-on agents called "dots" in September, and Meta released Muse, a personal agent that can shop, book travel, send email and make payments.49 A comparison table from Shattered lists Microsoft's Copilot agents and Amazon Q as generally available, while Google's agent is still in preview.11 Technology.org also points to a recent Microsoft Copilot rebuild that added an always-on Autopilot agent.5 On the same day, Anthropic added dashboard and animation tools to Claude.5

Google backed the launch with adoption figures. It says nearly 90% of the Fortune 100 use Gemini Enterprise, nearly 80% of Google Cloud customers use its AI products, and nearly 500 customers each processed more than a trillion tokens last year.3 These numbers come from Google and describe Gemini Enterprise as a whole, not the new agent.1017 Reported early testers include On, Shopify and PayPal.10 Technology.org cites a Gartner forecast that 40% of enterprise applications will include task-specific agents by the end of 2026, up from under 5% in 2025.5

Industry versions are part of the plan. Versions for financial services and legal work are in preview, and versions for government, health care and retail are coming.618

What is still missing

The gap between the announcement and what customers can actually use is large. The agent is in private preview. Google says wider availability is coming "soon" for select Workspace Business and Enterprise plans.1520 Several outlets point out that Google has given no price, no general-availability date and no confirmed list of eligible plans.1119 Some specifications in circulation, such as one outlet's claim of a 2-million-token context window, do not appear in the more widely sourced reports.13 Readers should treat them as unconfirmed.

There is also a practical change for existing users. Google is replacing Gems with a new Skills feature. Gems support for personal accounts ends in November 2026, and Workspace enterprise plans have until March 2027.16

Our read

This is a bundling strategy more than a technical leap. Google is betting that companies would rather run one agent with shared memory across Workspace, Slack and Microsoft 365 than manage separate tools for writing, meetings, notes and research.5 That poses a real threat to standalone AI productivity apps whose main advantage is convenience, because Google already holds the email, calendar and documents that give its agent context.

The open question is trust. Agents with their own identities and access to core business systems need security and reliability that IT departments can check.4 Without public pricing, availability dates or independent measures of how often tasks are completed, the launch is still an ambitious plan rather than a finished product.1719