Gemini Enterprise Agents: Agent Gallery, Agent Platform and Agentspace

Google's documentation defines it in one line: Gemini Enterprise is "an intranet search, AI assistant, and agentic platform" that "empowers knowledge workers with generative AI and agentic workflows by leveraging data sources from across your organization."

In practice it is three things stacked together. There is a permissions-aware search layer over your company data, which Google describes as "a single, multimodal search interface offering permissions-aware access to enterprise information." There is a chat assistant on top of that. And there is the agent layer, which is what this page is about: a gallery of ready-made agents, a builder for making your own, and the governance around both.

The data side runs on prebuilt connectors. Google documents "prebuilt connectors for commonly-used third-party applications" including Confluence, Jira, Microsoft SharePoint and ServiceNow, with over 100 platforms supported in total, covering Google Workspace, Microsoft 365, databases and CRMs. That connector list is the part that determines whether an agent can actually see the data it needs, which is why the edition gate further down matters more than the headline seat price.

What happened to Google Agentspace

Agentspace was the original name. Google renamed it to Gemini Enterprise on October 9, 2025, and the Agentspace name has since been retired from Google's materials. The functionality carried over rather than being replaced, so documentation and tutorials written about Agentspace generally still describe how Gemini Enterprise behaves. If you are reading a 2025 guide to Agentspace, the concepts hold, but the product name, the console and the edition structure have all moved on.

This is the single most confusing thing about the product right now, and it is confusing because Google gave two different products almost the same name four months ago. On April 22, 2026, at Next '26, Google announced the Gemini Enterprise Agent Platform, describing it as "our new developer platform and evolution of Vertex AI." Vertex AI, in other words, is being renamed.

So there are now two things called Gemini Enterprise something, and they serve different people.

Google frames the combined offering as "an end-to-end system for the agentic era, built for agents that can execute complex, multi-step work processes." The platform side is organized around four verbs. Build covers the Agent Development Kit, now with a graph-based framework and support for MCP. Scale covers the Agent Runtime, agent-to-agent orchestration, and Memory Bank for long-term contextual memory. Govern covers Agent Identity, which assigns "every agent a unique cryptographic ID for complete traceability and auditing," the Agent Gateway that acts as "air traffic control to ensure interactions between agents and data is secure," and Model Armor against "prompt injection, tool poisoning, and sensitive data leakage." Optimize covers Agent Simulation, to "stress-test your agents against real-world scenarios before they ship," and Agent Observability.

One practical note for anyone already on Vertex AI: the April announcement positions the Agent Platform as the evolution of Vertex AI but does not publish a migration plan or a deprecation timeline. If you have production workloads on Vertex AI, that timeline is a question for your account team rather than something you can read off a docs page today.

The Agent Gallery is where ready-made agents live inside the Gemini Enterprise app. It started as Google's own out-of-the-box agents. On April 23, 2026, Google opened it up: "Today, we're bringing partner-built agents from our Agent Marketplace directly into the Agent Gallery inside the Gemini Enterprise app. This creates a powerful, centrally governed hub for discovering and managing specialized AI from partners."

That announcement covered more than 80 partner-built agents. A sample of what is in there, and what each one actually does:

Others in the launch list include Accenture, Deloitte, Lovable, Palo Alto Networks, Replit, S&P Global and Workday.

The certification, and what it is worth

Google applies a quality bar rather than listing anything that shows up. In Google's words, "Every featured agent earns the 'Google Cloud Ready - Gemini Enterprise' designation only after passing a strict four-step evaluation for basic functionality, output accuracy, autonomous execution, and enterprise standards." Those four criteria are worth reading closely, because output accuracy and autonomous execution are the two that usually separate a demo from something you can leave running.

The approval queue

Discovering an agent is not the same as getting it. Google is explicit about the workflow: "Employees can browse the gallery and submit procurement or access requests for the agents that fit their needs, while IT remains firmly in the driver's seat. Administrators review all incoming requests to approve or deny deployments, retaining granular control over exactly who can interact with each agent across the organization."

That is the right design for a large regulated organization and a genuine obstacle for a small team that just wants to start. It is the same pattern as the Microsoft Agent Store, where every inclusion route also ends at an admin approval step. If you are a five-person team without a Google Cloud admin on staff, budget for the queue as well as the licence.

Agent Designer is the no-code path. Google defines it as "an interactive no-code, low-code platform for creating, managing, and launching single and multi-step agents in Gemini Enterprise." It is aimed squarely at people who are not going to open the Agent Development Kit.

The interface has two panes. The chat pane is conversational: you describe the agent you want in natural language and it gets created and previewed. The designer pane is the low-code half, with three tabs, Flow for the workflow itself, Schedule for when it runs, and Preview for testing before you ship.

Agents built this way can connect to Google and third-party data sources including Gmail, Google Drive and Jira, and can be scheduled to run on a recurring basis rather than only on demand. Scheduling is the feature that turns a helper into something closer to a background worker, so it is worth checking early whether the job you have in mind fits a schedule or needs an event trigger.

What the documentation does not currently spell out is the publishing and sharing model, usage limits, or prerequisites beyond general Gemini Enterprise access. If your plan depends on one team building an agent and twenty teams consuming it, confirm how that sharing works before you design around it.

This is the section to read twice, because the edition that blocks the most is the one that looks safest to start on. Every figure below is read straight off Google's published edition comparison, checked again on August 20, 2026. The quoted strings are Google's own wording.

The Business edition cannot reach the Agent Marketplace

Read the table column by column and the gate falls out in one line. Business is the only edition in Google's comparison with no check against "Access the Agent Marketplace." It is also the only edition that does not "Access the full data connector ecosystem," the only one without "Use Gemini Code Assist Standard," the only one without "Use Data Insights (Made by Google)," and the only one with no check against "Enterprise-grade security and compliance." Standard, Plus, Pay-as-you-go and Frontline all carry the marketplace and full-connector checks.

That matters more than the seat price, because installing agents somebody else built is usually the reason a company buys this product in the first place. Business is capped at 500 users and priced as the small-organization entry point, so the buyers most likely to pick it are the ones least able to build agents themselves, and it is the edition that shuts the catalog. If partner agents are on your requirements list, Standard is the floor, not Business. List the systems your agents must read and the agents you expect to install, then confirm both against the specific edition on your quote before you sign.

One more oddity worth knowing: Frontline, the cheapest seat in the lineup at 2 GiB of storage, does carry both the full connector ecosystem and Agent Marketplace checks. It is restricted in other ways, covered below, but connector reach is not one of them.

Frontline is an add-on, not an entry point

Frontline looks like the cheap way in and is not available as one. It requires "150 or more users of Standard or Plus editions" already in place, ships 2 GiB of pooled storage, excludes Gemini Code Assist, and limits users to agents someone else already built and provisioned for them. It is a way to extend an existing deployment to deskless staff, not a way to start small.

What the seat price leaves out

Google does not publish a complete public price list for Gemini Enterprise, so seat figures in circulation are reseller-reported list prices: roughly $21 for Business, $30 for Standard on a 12-month commitment, and $50 for Plus on an annual commitment, with higher numbers without commitment. Treat those as a starting point for your own quote rather than as published prices. Compute consumption also bills separately to your linked Google Cloud account rather than being included in the seat. The full breakdown, including what moves the number, is in our Gemini Enterprise pricing analysis.

Governance is the strongest part of the offering and the reason large organizations pick it. It works at two levels.

Inside the app, the control is the approval queue described earlier: staff request, administrators approve or deny, and admins retain granular control over exactly who can interact with each agent. Nothing gets deployed because an employee liked the look of it.

On the platform side, the controls are lower level and more interesting. Agent Identity assigns "every agent a unique cryptographic ID for complete traceability and auditing," which is the same architectural idea Microsoft implemented as Entra Agent ID. The Agent Gateway sits between agents and data as "air traffic control to ensure interactions between agents and data is secure." Model Armor defends against "prompt injection, tool poisoning, and sensitive data leakage," three attack classes that are specific to agents rather than inherited from ordinary application security.

Agent Simulation is the one that gets least attention and deserves more. Being able to "stress-test your agents against real-world scenarios before they ship" is the difference between finding out an agent mishandles an edge case in staging and finding out from a customer. If you are evaluating platforms, ask what the equivalent is on any competitor, because most do not have one.

Access is permissions-aware throughout, meaning search and agents respect the underlying document permissions rather than flattening them. That property is what makes enterprise search safe to switch on, and it is worth verifying during a pilot with a deliberately restricted test account rather than taking on trust.

Whatever platform you land on, the governance questions are the same ones. Our guide to enterprise AI agent security covers the checks worth running before an agent gets access to production data.

Five things to weigh before committing, all of them from Google's own materials or the absence of them.

1. The naming is genuinely unstable. Three names in under two years, and the current two differ by two words. Internal documentation, procurement records and training material all need to say which product they mean. Expect to re-explain this to your own colleagues more than once.
2. There is no published Vertex AI migration path. The April 2026 announcement calls the Agent Platform the evolution of Vertex AI without publishing a transition plan or deprecation timeline. Existing Vertex AI workloads need that answered by your account team.
3. Business is gated out of the marketplace. Covered above, and worth restating because it is the most expensive misread available here: the entry edition is the one that cannot install partner agents.
4. Pricing is not fully public. Google publishes edition capabilities but not a complete price list, so every figure you find is reseller-reported and every serious number comes from a quote.
5. Agent Designer sharing is undocumented. Publishing, sharing and governance workflows for agents you build yourself are not spelled out in the current documentation, so validate them in a pilot rather than assuming.

When Gemini Enterprise is the wrong tool

It is a strong fit for a large organization already on Google Cloud that wants permissions-aware search over company data, a governed catalog of agents, and a no-code builder for internal teams. It is a poor fit in three situations.

The first is when you are small. Frontline requires 150 Standard or Plus seats before it exists, the approval workflow assumes a real IT function, and the platform overhead is priced for organizations that need it. The second is when your stack is not Google-shaped. The connector list is broad, but an agent whose job lives in a system that is not on it will not reach the data regardless of edition. The third, and most common, is when you want an agent rather than an agent platform. Buying Gemini Enterprise to get one working agent is buying a platform seat, a connector configuration and an approval queue to reach something you could have deployed directly.

That last case is what Agentmarketplace is for. You browse ready-made agents by category, see the integrations and permissions each one needs before you commit, and deploy onto the stack you already run without a platform seat underneath it. For how the vendor catalogs compare against each other, see AI agent marketplaces, and for the closest Microsoft equivalent to everything on this page, Microsoft Agent 365.