On Oct. 8, 2026, Google Cloud used its "Gemini at Work 2026" event to unveil Gemini Agent, an enterprise AI that does not always run on its maker's own models. Google says the agent automatically selects between its Gemini model family and Anthropic's Claude for each job, with additional proprietary and open-source models planned. Google Cloud CEO Thomas Kurian outlined the product in a blog post adapted from his keynote.
The announcement itself is confirmed, but important commercial details are not. Google has not disclosed separate pricing, a general release date, or independent benchmarks, and the agent remains in private preview, according to Google Cloud's post and reporting from Yellow and Zeniteq.
An agent that takes objectives, not instructions
Kurian's framing is that users give the agent "objectives, not instructions." Rather than following step-by-step commands, it plans tasks, calls tools, connects to internal business systems, and returns finished work. Inside Google Workspace, it operates inline in Gmail, Drive, Docs, Slides, Sheets, Chat, and Calendar, carrying the same memory and skills across surfaces. It can also be reached from Slack, Microsoft 365, and the command line, and it connects to external systems including Salesforce, ServiceNow, Snowflake, Jira, Confluence, Git, BigQuery, Databricks, Postgres, and MCP servers.
For longer jobs, the agent can spin up temporary sub-agents that work in parallel or sequence, potentially for hours or days. A separate "coworker" mode gives an agent a persistent role with its own Workspace account, email address, calendar, Drive storage, and presence in the company directory — while limiting its access to what a team shares with it. Google says the agent keeps four kinds of memory — session, semantic, procedural, and episodic — so it accumulates knowledge, procedures, and experience over time.
Why Google would hand tasks to Claude
Claude's appearance inside Gemini is less novel than it looks. Anthropic's models have been available on Google Cloud since March 2024, when the Claude 3 series arrived on Vertex AI, and the two companies are deeply entangled on compute: in October 2025 Anthropic announced it would use up to 1 million Google TPUs in a deal worth tens of billions of dollars, with more than a gigawatt of capacity coming online in 2026.
What changed is position. Claude previously sat on a shelf for developers to pick and wire up; now it is a back-end resource inside Google's agent. Google's stated rationale is that leading models change every few months, and that keeping the model layer open lets enterprises swap models without rebuilding the surrounding context, skills, and tool connections. As one analysis of the launch put it, opening the model layer means customers do not leave for a better model, while memory, skills, and the working environment stay inside Gemini — forming a hard-to-migrate retention moat.
That logic reflects a wider shift in the industry. As model capability has become increasingly commoditized — with the gap between top models narrowing on benchmarks — vendors argue that contextual access, not raw intelligence, is becoming the real differentiator. A related industry discussion frames the developer skill set as moving from prompt engineering toward "context engineering," a system-level methodology for managing dynamic, structured, multi-source information flows as the AI's working memory.
Governance and cost as the enterprise pitch
Kurian built the announcement around two questions he says determine whether enterprise agent programs succeed or stall: whether a company can govern them, and whether it can afford them. Every agent receives its own identity, its actions land in an audit trail attributed to the agent rather than a person, and real-time spending caps pause a project's agent when a limit is hit. Google also describes an Agent Sandbox and an Agent Gateway that enforces organization-wide policies, such as blocking access to documents in a given confidentiality class.
Google's scale figures describe its existing AI products, not the new agent specifically. The company says nearly 80% of Google Cloud customers use its AI products, nearly 90% of the Fortune 100 use Gemini Enterprise, and almost 500 customers each processed more than one trillion tokens over the past year. Named early testers include On, Shopify, and PayPal, with customers including BNP Paribas, Bradesco, Merck, Orange Spain, Santee Cooper, SOMPO, Ulta Beauty, and Wesfarmers. Google reports that Bradesco cut document review time from one hour to five minutes and that Orange Spain has deployed more than 1,000 custom Gemini Enterprise agents.
Anthropic is moving in parallel
The announcement lands in a crowded field. Microsoft added a persistent Autopilot agent to Copilot on Sept. 25, and OpenAI introduced its always-on Dots agents days later. Anthropic itself brought Claude directly into Workspace's Docs, Sheets, and Slides on Oct. 6 — two days before Google's event — and has been testing Claude Tag in Slack, a permanent channel-level agent with persistent memory and an "ambient" mode that lets it act without being prompted, with automatic workspace migration planned to begin Aug. 3, 2026.
Anthropic's commercial momentum is also unusual for a rival embedded inside Google's product. The company closed a financing round at a $965 billion valuation, surpassing OpenAI, and said it had confidentially filed for an IPO, CNBC reported. Its annual recurring revenue rose from roughly $1 billion at the end of 2024 to about $45 billion by May 2026, per a Huxiu analysis. Anthropic also unveiled Claude Opus 4.8 with a 1 million-token default context window.
The two companies represent sharply different strategic paths, according to an analysis by Small Island Research Notes: Google pursues a full-platform route — models, cloud infrastructure, enterprise tools, and global scale — absorbing AI into existing systems, while Anthropic pursues a focused route built on model capability, enterprise trust, and cross-platform positioning.
What is confirmed, unknown, and awaited
Confirmed: Gemini Agent was announced Oct. 8, 2026; it routes tasks between Gemini and Claude models with more models planned; it works inside Workspace and is reachable from Slack and Microsoft 365; it supports coworker agents with dedicated identities and four types of memory; and it ships with governance controls including audit trails and spend caps.
Unknown: Pricing, a general release date, and independent benchmarks have not been disclosed. Google has not published its full routing criteria or evidence that multi-model selection improves quality or cost in controlled comparisons. Adoption figures cover Google's broader AI products rather than the new agent, and many announced capabilities remain in preview with per-feature availability varying by plan and region.
Awaited: A public rollout timeline and pricing, third-party evaluations of task completion rates and total cost, and details on how memory is retained, updated, and corrected as personnel and procedures change. Also unresolved is how durable the model-layer openness is — and whether enterprises will treat the accumulating context, skills, and connections as reason enough to stay.