Gemini agent is Google’s unified enterprise AI agent, introduced on October 8, 2026: you give it a goal, not a prompt, it plans the work on its own, connects to workplace systems, can be assigned its own email account, and records every action in a log in its own name. For a Russian company, the product itself is most likely unavailable, but its architecture is a useful ready-made checklist. Six Google design choices — goals instead of commands, a separate agent account, skills, memory, a least-privilege log, and strict spending limits — can be replicated on your own stack: with 1C, Bitrix24, corporate email, and a model deployed in Russia.
Prepared by AIrassvet using open sources, verified on October 11, 2026. It is intended for business owners and executives, IT directors, and digital transformation leaders who are deciding what an AI agent should look like in their company and how to control it. We did not test Gemini agent: the product description is based on a Google publication and coverage by industry outlets. AIrassvet is not affiliated with Google, SAP, or any other companies mentioned, and the section at the end describes the publisher’s services.
Contents
- What Google announced on October 8, 2026
- In brief: what matters for business
- What a universal work agent is
- Six design choices in the Gemini agent architecture
- What Google did not say
- SAP Joule Work: the same idea from the system of record side
- What this means for Russian companies
- How to build a digital employee agent on your own stack
- What to implement now, and what later
- Risks of an agent with its own account
- Frequently asked questions
- Primary sources and verification method
What Google announced on October 8, 2026
At Gemini at Work 2026, Google Cloud CEO Thomas Kurian introduced Gemini agent. In the Google Cloud announcement , it is described as a single universal work agent that answers questions, performs knowledge work, creates images and media, writes and runs code — through one interface and one API.
Key facts from the source:
- Goals instead of instructions. The agent plans the work, uses skills and tools, connects to company systems, and returns a finished result. Tasks can be assigned manually, on a schedule, or by an event.
- Works wherever the employee works. Web, iOS, Android, Windows, Mac, command line, Google Workspace, Microsoft 365, Slack, and third-party apps, plus a no-UI mode.
- Long-running tasks in the cloud. Work that takes hours or days keeps running after the laptop is closed.
- Agent coworkers. An agent with a persistent role gets its own account with an address like @agents.company.com, its own storage, calendar, and place in the employee directory. For parallel steps, it creates temporary subagents, each with its own identity.
- Model selection. The agent distributes work between Gemini models and Anthropic Claude models, and in the future, other models as well.
- Connections. Confluence, Microsoft Office, Teams, Slack, Workspace, Git, Jira, Salesforce, ServiceNow, BigQuery, Databricks, Postgres, Snowflake, local files, and any MCP server inside or outside the company network.
- Industry-specific versions. For finance and legal tasks — early access; for the public sector, healthcare, and retail — later.
According to TechCrunch, Google is launching the agent for business first, with consumer access coming later. Sundar Pichai said at the event that Gemini has more than 1 billion monthly active users, and nearly 90% of Fortune 100 companies use Gemini Enterprise. These are company claims, not independent data.
In brief: what matters for business
- The agent stops being a chat interface. It is assigned an outcome, not questions to answer; it works for hours and decides on its own what steps to take.
- The agent gets a separate account. Actions are logged in the agent’s name, not a person’s, and are visible in the audit trail — this changes the approach to access rights and accountability.
- Google is moving control to the infrastructure layer. Isolated environment, network gateway with policies, least privilege, and a spending cap that stops the agent when reached.
- There is no public price or general availability date. Google’s publication does not include them; this is an announcement of direction, not a finished pricing plan.
- The trend is confirmed by more than Google. Two days earlier, SAP showed Joule Work — an interface where an employee states an intent, and agents do the work in the system of record.
- A Russian company does not need to wait for a Google product. The architectural choices can be replicated on your own systems, starting with one agent and one process.
What a universal work agent is
A universal work agent is one AI agent that employees assign a variety of tasks to through familiar channels, while it independently chooses tools, data, and models, performs multi-step work, and reports back with the result. Its opposite is a collection of narrow bots, each of which can do only one action in one system.
The difference from a typical AI assistant comes down to three things. An assistant responds within a single conversation, while an agent carries a task through to completion. An assistant acts on behalf of the user, while a coworker agent acts in its own name and with its own permissions. The user controls an assistant; an administrator controls an agent through policies, logs, and limits.
The idea is not new: we have previously written in detail about an AI agent as a digital employee. What is new in Google’s announcement is that a major vendor has made this model a standard part of an enterprise product.
Six design choices in the Gemini agent architecture
We distilled Google’s description into six design choices that determine how the agent works and how it is controlled. In the right-hand column is how to replicate the choice without Google’s product.
| Design choice | How Google does it | Why it matters | How to do it yourself |
|---|---|---|---|
| 1. Goals instead of instructions | The task is framed as an outcome; launch can be manual, scheduled, or event-driven | The agent plans the steps itself, and a person approves the result | Define common agent tasks with completion criteria and triggers: email, new deal, time |
| 2. Its own account | A coworker agent with email @agents.company.com, storage, and calendar; subagents with their own identities | Actions are separated from human actions, and permissions are configured separately | A service account for the agent in email, 1C, and CRM with a separate role |
| 3. Skills and a tool registry | Skills are modular instructions and processes; a shared registry of skills and tools; MCP | Repeatability: a process defined once works for everyone | A skill library in the knowledge base, tools exposed through MCP servers for 1C, CRM, and email |
| 4. Memory | Session, semantic, procedural, and episodic memory | The agent does not start from scratch every time | Project- and client-specific context storage with access controls |
| 5. Governance and audit | Verified identity, least privilege, security-assigned roles, actions logged under the agent’s name, isolated environment, gateway filter | It is clear who the agent is, what it is allowed to do, what it did, and where it cannot go | Least-privilege role, full action logging, test environment, restricted data list |
| 6. Cost control | Model selection by task, automatic routing, hard project spending cap with agent shutdown | The agent does not burn through the budget endlessly | Limits at the model gateway level, routing simple tasks to a low-cost model |
A few details worth adopting verbatim.
The agent only sees what has been shared with it. Google says the colleague agent operates under its own identity and only sees what it has been granted access to. That is the opposite of the common setup where an agent runs under a manager’s account and can see everything.
Policies are defined explicitly. As an example, Google cites a rule that agents cannot open documents marked restricted. In your own setup, define these rules before launch: which folders, document types, and fields the agent must never open.
A spending limit stops the work. When the project agent reaches the threshold, it is paused, and work resumes with one click. Spending is tracked by project, so it can be allocated across departments.
Saved reports do not consume tokens. Google notes that a report generated by an agent can later be run on demand without model costs. The principle is general: if the agent wrote a data request once, you can save it as a standard report and avoid generating it again.
What Google did not say
Google’s publication does not include:
- pricing per user or per token for the Gemini agent itself;
- a general availability date for the main agent — preview access is listed only for industry-specific versions;
- a list of regions and editions where the agent will be available;
- details on data handling beyond the described logging and isolation.
The named customers confirm scale, not results for your company. For example, according to Google, PayPal routes 10 million requests to different models per week, and Orange Spain has more than 1,000 in-house agents. The methodology behind these figures was not disclosed.
SAP Joule Work: the same idea from the ERP side
On October 6, 2026, at SAP Connect, SAP introduced Joule Work. According to ERP Today, it is a natural-language interface through which employees ask questions and carry out work in SAP data and third-party systems; answers are based on SAP’s knowledge graph. Along with it, SAP announced new assistants for supply chains, foreign trade, customer service, and finance.
The timeline is stretched out: some capabilities will arrive in Q4 2026, customer service and spend-analysis assistants in Q1 2027, and new finance assistants will be available only to early adopters in Q1 2027.
Why this matters alongside Google’s announcement. Google is coming from the workplace side — email, documents, and calendar. SAP is coming from the ERP side, where orders, postings, and inventory balances live. Both companies are moving toward the same pattern: the employee states what they need, and the agents gather the data and execute actions in the systems. For a Russian company, the ERP-system equivalent is most often 1C, and that is where an agent brings the most value and carries the most risk.
What this means for Russian companies
In August 2026, according to the media, Gemini stopped opening for users in Russia, including via VPN. There are no official access terms for Gemini agent for Russian companies, and building workflows around circumvention is risky: access can disappear along with settings and history. We looked at a similar situation with Claude in the article about the Claude block in Russia.
The practical takeaway is different: the value of the announcement is in the architecture, not the specific product. Google spent resources answering the four questions that come up for any company using agents: who the agent is, what it is allowed to do, what it did, and what it must never touch. These questions do not depend on the model vendor.
For a Russian company, such an agent needs:
- a modelthat is legally accessible: a Russian or open model on your own servers or in a Russian cloud;
- connections to your systems: 1C, Bitrix24 or amoCRM, corporate email, messengers;
- the agent account with a separate role in each system;
- logging and limitsthat do not depend on the model itself.
How to build an employee agent on your own stack
The approach we recommend follows Google’s logic, but starts with one process rather than a universal agent for everything.
- Choose a role, not a technology. For example, “purchasing manager assistant”: collects requests from email, checks inventory in 1C, and prepares a draft order for the supplier. The role defines both the tasks and the limits of access.
- Describe tasks as outcomes. For each task, define the input data, expected result, completion criteria, and what counts as an error. These are the agent’s future skills.
- Create the agent account. A separate mailbox, a separate user in 1C and CRM with a role that allows only what is necessary. The agent should not work under an employee’s account.
- Connect systems through tools. MCP servers or APIs for 1C, CRM, and email, where each tool performs one action and checks permissions. We described how to connect agents to an ERP system in the article about connecting AI agents to 1C, and for CRM in the piece on an AI agent for Bitrix24 and amoCRM.
- Separate read and write actions. Reading — without confirmation. Creating a draft — with notification. Sending to the supplier, posting the document, payment — only after human approval. For more on levels of autonomy, see the article on managing AI agent autonomy.
- Turn on logging before the first launch. Every action: time, task, tool, input data, result, and who authorized it. This is analogous to Google’s log, where actions are attributed to the agent.
- Set limits. A monthly model budget with agent shutdown on overrun, a cap on the number of steps in a single task, and routing simple steps to a low-cost model. For routing, see the article on AI model routers.
- Test it in a sandbox. Run standard and edge-case tasks on a copy of the data and test accounts, as described in the guide to an AI agent sandbox.
- Expand the role gradually. Add new tasks and permissions only after logs over several weeks show that the agent handles the current ones well.
Why we recommend starting with one role instead of a universal agent like Google’s. A universal agent requires mature infrastructure: a skills registry, a unified permissions model, and a centralized log. At Google, that’s part of the platform; for most companies, it isn’t. One role lets you build these elements at a smaller scale and test them. The tradeoff is that, at the first stage, the agent handles fewer tasks, and the benefit is visible in one department rather than across the entire company.
What to implement now and what to postpone
| Element | Now | Later | Why |
|---|---|---|---|
| Separate agent account | Yes | — | Without it, you can’t separate permissions or investigate incidents |
| Activity log | Yes | — | Needed from day one, even for a single agent |
| Model spending cap | Yes | — | Cheap to set up, expensive not to |
| Reading data from 1C and CRM | Yes | — | Major benefit with relatively low risk |
| Writing to the accounting system | After testing | Yes | An accounting error is more costly than a reporting error |
| Sending emails to customers on the agent’s behalf | No | After testing | Reputational risk and the risk of following someone else’s instructions from emails |
| Long-term customer memory | Limited | Yes | Rules are needed for storing personal data |
| Universal agent for the whole company | No | Yes | Requires a skills registry and a unified permissions model |
Risks of an agent with its own account
An agent-colleague has risks that a chat assistant doesn’t:
- A new account is a new attack surface. If the agent can be asked to do something by email, an attacker can email it too. The agent should treat instructions in incoming documents as data, not commands. Common attacks are covered in the article on prompt injection and attacks on agent tools.
- Permissions expand unnoticed. Every separate “give the agent access to…” request seems harmless on its own. The agent’s permissions should be reviewed on a schedule, just like employee access rights.
- Accountability gets blurred. If the agent acts on its own behalf, you need to define in advance which people are responsible for its actions in each process.
- Costs are unpredictable. An agent that plans its own steps can get stuck in loops. That’s why Google puts the spending cap at the platform level.
- Agents multiply uncontrollably. If any employee can create agents, in six months the company may have dozens of agents with unknown permissions. You need an agent registry: who created it, what role it has, what permissions it has, and who is responsible. We described how to monitor agent work in the article on monitoring LLMs and AI agents.
Frequently asked questions
What is Gemini agent?
It is Google’s enterprise AI agent, introduced on October 8, 2026. You give it a goal, and it plans the work on its own, uses skills and tools, connects to company systems, chooses a model for the task, and returns a finished result. The agent can get its own account with email.
How much does Gemini agent cost?
Google did not specify pricing for the agent itself in its October 8 announcement. It described only spending-control mechanisms: model selection by task, automatic routing, and a project spending cap that stops the agent once it is reached.
Is Gemini agent available in Russia?
There are no official terms for Russian companies, and in August 2026, according to media reports, Gemini stopped opening for users in Russia. For business workflows, it’s better to build a similar agent on legally available models and your own systems.
How is an agent-colleague different from an AI assistant?
An assistant responds within the conversation and acts on behalf of the user. An agent-colleague carries the task through to completion, works for hours, acts on its own behalf with its own permissions, and its actions are logged and monitored by an administrator.
Can you build this kind of agent for 1C and Bitrix24?
Yes. You need a legally available model, connections to 1C, CRM, and email through APIs or MCP servers, a separate agent account with minimal permissions, an activity log, and spending limits. It’s best to start with one role and one process.
What is SAP Joule Work?
It is a natural-language interface introduced by SAP on October 6, 2026, through which employees ask questions and do work in SAP data and third-party systems. New assistants are being rolled out gradually in Q4 2026 and Q1 2027.
Original sources and verification method
We read the original sources on October 11, 2026. The description of Gemini agent is taken from Google Cloud’s post based on Thomas Kurian’s presentation; the user-count claims are from TechCrunch’s coverage; the details on SAP Joule Work are from ERP Today, because we could not find SAP’s October 6 press release at the time of verification. The six decisions, the “now or later” table, and the rollout order are AIrassvet’s editorial methodology, not the result of product testing.
- Google Cloud, 08.10.2026: Welcome to Gemini at Work 2026: Introducing the Gemini agent — capabilities, identity, skills, memory, governance, costs, customers.
- TechCrunch, 08.10.2026: Google brings agentic AI to Gemini, starting with businesses — rollout order, Sundar Pichai’s statements.
- ERP Today, 08.10.2026: At SAP Connect, SAP expands Joule across every line of business — Joule Work and the assistant rollout timeline.
- Mentoday, 14.08.2026: Gemini stopped working for Russians even with VPN — reports about access restrictions.
Restrictions: this is an announcement, and the product’s capabilities, timing, and terms may change; customer and user numbers are company claims without a disclosed methodology; information about availability in Russia is based on media reports, and we did not find an official statement from Google.
How AIrassvet helps build an AI employee agent
If you want the agent to operate in your systems under its own account and with clearly defined permissions, it’s better to design the architecture before the first connection. On the AIrassvet website the following directions are described, and they relate directly to this task:
- Development of AI agents integrated with the company’s communication channels and business systems.
- Agentic automation of actions in ERP, 1C, and browser interfaces.
- Deployment of a language model on the company’s own servers when data cannot be transferred to an external service.
- Process audit before implementation, integration, testing, launch, and ongoing support of the agent.
The first step is to choose one role for the agent, describe the process, its current metrics, systems and data, constraints, and the criterion by which you will consider the agent ready for work.
Bottom line: the agent becomes an employee, and it needs to be managed like one
The Gemini agent announcement shows where enterprise agents are headed: from chatbot assistant to digital employee with goals, an account, skills, memory, and logs. SAP Joule Work, announced the same week, confirms that this is not a one-vendor idea. But as the agent becomes more autonomous, the amount of control also grows: Google is building isolation, a network gateway, minimum permissions, and a strict spending cap around the agent.
A Russian company does not need to wait for access to Google’s product. Start with one role: a separate account, minimum permissions, logs, a spending cap, and a test environment. When that agent runs for several weeks without surprises, you have a foundation on which you can expand the role and add new agents.