Gemini Agent Launches With MCP Tools and Claude Model Routing

Google turns Gemini from an assistant into a worker

On Thursday, at its Gemini at Work 2026 event, Google Cloud introduced what it calls the "Gemini agent." The company describes it as one "universal agent for work" that customers can reach through the Gemini Enterprise app, Google Workspace and third-party services.2 Google Cloud CEO Thomas Kurian announced the product in a keynote. His pitch was a coworker that takes on assignments by itself and hands back finished work, not a chatbot that waits for the next prompt.3 Google's own slogan for this is that users should "give it objectives, not instructions."2

The list of capabilities is broad. The agent answers questions, does knowledge work, creates images and other media, and writes and runs code.23 It works inside Gmail, Docs, Sheets, Slides, Chat and Calendar. It also reaches into Slack and Microsoft 365, and runs on the web, iOS, Android, Windows and Mac.45

Outlets describe the launch in very similar terms, which is unsurprising because most of them are working from the same Google blog post. They differ in emphasis. General-news coverage treats it as Google's entry into a crowded race among tech giants for virtual office assistants.14 Trade and technical coverage focuses on the architecture. Taken together, that architecture is the story for enterprise buyers: sub-agents, persistent memory, model choice, and a tool layer built on the Model Context Protocol (MCP).

Long-running, cloud-resident and multi-agent

The agent runs in Google's cloud, not on the employee's computer. Google's argument is that this lets a job continue for hours or days after the user closes the laptop.48 Because the agent lives in the cloud, it keeps one set of memories, context and personalization across every device.2 Google says it has four kinds of memory. Session memory covers the current task. Semantic memory is a knowledge base the agent builds as it reads documents and talks to people. Procedural memory holds the skills it writes for itself. Episodic memory is a record of everything it has done before.2 Kurian summed up the benefit by saying users never have to re-brief it.3

For larger jobs, the agent creates "sub-agents." Google describes these as temporary, task-specific agents, each with its own identity, that can work in parallel or one after another over several days.34 There is also a longer-lived category Google calls "coworker agents." These are digital team members with their own Workspace accounts, email addresses, calendars and Drive storage. They can be added to a Chat space or mentioned in a document.5 One detailed write-up says these accounts sit under an @agents.company.com address and can see only the context a user or team gives them.9

Google's examples are ordinary office work. A manager emails asking for a project update as a slide deck, and Workspace offers a one-click option to hand the job to Gemini. A user asks it to schedule a meeting with "the usual team of regional event leads" without naming anyone, and it works out who they are.2 In that second case, the agent figures out the attendees from chat-space membership and past threads, checks their calendars, and starts the email exchange, even when outside guests are involved.8

MCP is the plumbing

The MCP angle gets the least attention in general coverage, but it is where Google's strategy is easiest to read. In its launch post, Google says the agent can securely connect to any MCP server, whether inside or outside a company's network. It also offers an enterprise tools registry so teams can build tools and publish them across the company.19 Named integrations include Salesforce, ServiceNow, Jira, Git, BigQuery, Databricks, Postgres and Snowflake, and companies can add their own tools and reusable skills.8

That support didn't arrive all at once. Gemini Enterprise's release notes show the groundwork went in over several months. Custom MCP server data stores appeared in April and reached general availability in August. Agent Registry, a catalog of agents and MCP servers, came to Gemini Enterprise in June. In early October, a federated query mode arrived that uses MCP and each user's own credentials to query data where it already sits, without copying it.14 The Agent Platform release notes fill in the rest. Agent Identity became generally available in April, Agent Registry reached GA in June, and IAM Unified Access Policies for controlling which MCP servers and agents an agent can talk to reached GA at the end of August.15

One point stands out. Google has built a policy layer on top of MCP traffic. Its Agent Gateway documentation says the gateway can parse MCP requests and enforce rules down to the level of individual tools. It names MCP prompt injection as a specific threat, to be countered with Model Armor.13 A Google codelab goes further, describing direct network access for autonomous agents as a data-exfiltration risk. It shows how to route Gemini Enterprise's calls to private MCP servers through the gateway, with each call carrying a verifiable agent identity.12 Google's remote MCP servers have also moved to the stateless version of the protocol dated 2026-07-28, which removes the session handshake in favor of self-describing HTTP requests.11

The reading here is straightforward. Google is treating MCP as a basic standard, as ordinary as HTTP, and competing on governance rather than on owning the connectors. That fits how others describe the protocol: model-agnostic, so the same servers work with Claude, OpenAI or Gemini hosts.17

Model choice, including a rival's

The most unusual design choice is that Google's agent is not tied to Google's models. It sends each job to whichever model suits it best. Today that means the Gemini and Claude families, with other private and open models promised later. Google pitches this as a way to improve quality and cut costs.27 One detailed analysis quotes Kurian arguing that the best model for a task is not always the largest, and that the leader changes every few months. It also cites a Constellation Research analyst who called the open approach the standout feature, because enterprises can bring models they already trust and pay for.9

This is a calculated concession. Google is putting its value in the agent layer, meaning memory, identity, tools and governance, and treating the model as something customers can swap out. The MCP strategy follows the same logic. If both tools and models are interchangeable, Google's lock-in has to come from the orchestration and the Workspace data underneath it.

Governance, cost and the competition

Google stresses enterprise controls throughout: administration, authorization, permissions and sandboxing.4 According to one account, every agent runs tasks inside an Agent Sandbox with its own network boundary, and all traffic passes through Agent Gateway, where a single policy can apply across the whole company.9 The same account describes cost features, including automatic "Smart Routing" to cheaper models and deferred off-peak execution that Google says can cut inference costs by up to half for eligible jobs.9

Google is also leaning on its installed base. The company says nearly 80% of Google Cloud customers use its AI products, nearly 90% of the Fortune 100 use Gemini Enterprise, and close to 500 customers each processed more than a trillion tokens last year.4 Versions for financial services and legal work are in preview, with government, health care and retail versions planned.410

The competitive context is crowded. Reports place the launch alongside OpenAI's "Dots" agents and Meta's "Muse."410 One comparison also lists Anthropic's Claude for Workspace, announced two days earlier, and a revamped Microsoft Copilot from late September.9

What remains unproven

The biggest gap is availability. The same analysis reports, citing 9to5Google, that the agent is in private preview. Wider availability is promised "soon" for some Business and Enterprise Workspace plans, with no general availability date and no published price.9 Several stories present the product as fully launched, which overstates where it actually stands.13

Reliability is the second open question. Google has not offered independent benchmarks for long, cross-application jobs, and an agent that runs for days can let a small early error grow.9 The appeal will depend on whether the agent can reliably finish useful work under real corporate oversight.5

In short: the agent's capability list is no longer what sets products apart, since every major lab now claims to write code and run tasks. Google's real bet is that enterprises will choose the vendor whose agents are easiest to govern. That means verifiable identities, MCP traffic that can be filtered tool by tool, and models that can be swapped. The infrastructure for that bet shipped over the past six months. Whether the agent built on top of it performs as advertised can only be judged once it leaves preview.