Google's Gemini agents can now take a goal, plan the steps, work across your apps and business systems, and run code in Google-hosted sandboxes. Four products do it: the Gemini agent for business (announced 8 October 2026), Gemini Spark for individuals, Managed Agents in the Gemini API for developers, and Gemini Enterprise Agent Platform, which is what Vertex AI became. They differ in who can use them, what they cost and how much you build yourself.
I read Google's release notes, changelogs and announcements for every date here. I run Gravity, an agent platform in public beta whose agents run on model APIs, so I watch these releases closely. Most coverage dates from Google I/O in May (our Google I/O 2026 recap covers that moment). A lot shipped after it.
What changed in 2026: the dated timeline
Google shipped something for agent builders in almost every month of 2026. Dates come from the Gemini API changelog, the Agent Platform and Gemini Enterprise release notes, and Google's posts on Spark and Gemini at Work.
| Date | What shipped | Who it is for |
|---|---|---|
| 11 December 2025 | Interactions API and the Deep Research agent, in preview | Developers |
| 18 March 2026 | Built-in tools and custom functions in one API call | Developers |
| 22 April 2026 | Vertex AI becomes part of Gemini Enterprise Agent Platform; Agent Engine becomes Agent Runtime; Agent Identity GA | Platform teams |
| 19 May 2026 | Managed Agents and the Antigravity Agent in public preview; Gemini 3.5 Flash GA | Developers |
| 19 May 2026 | Gemini Spark announced; Daily Brief starts rolling out | Individuals |
| 18 June 2026 | Agent Gateway, Observability and Registry GA | Platform teams |
| 24 June 2026 | Computer Use in public preview on Gemini 3.5 Flash | Developers |
| 21 July 2026 | Gemini 3.6 Flash and 3.5 Flash-Lite GA | Developers |
| 2 September 2026 | Gemini 3.8 Flash GA; deferred tier at a 50% token discount (Preview) | Developers, platform teams |
| 3 September 2026 | Workflow Builder GA in Gemini Enterprise | Business teams |
| 9 September 2026 | Computer Use and Shell sandboxes GA on Agent Platform | Platform teams |
| 17 September 2026 | antigravity-preview-09-2026 replaces the May version | Developers |
| 5 October 2026 | antigravity-preview-05-2026 shuts down | Developers |
| 8 October 2026 | The Gemini agent announced; admins can switch on Claude Opus 5.5 and Sonnet 5.5 in Gemini Enterprise's Antigravity tools | Businesses |
Most rows serve developers and platform teams. Individuals got one launch; businesses got Workflow Builder in September and the headline agent in October. Names keep moving too, so check the date on any tutorial you follow.
Four Gemini agents, compared
Prices are US list prices from Google's subscriptions page, and Spark's availability is from its product page. All of it is as of 9 October 2026.
| Product | Who it is for | What it does | Availability, 9 Oct 2026 | Price |
|---|---|---|---|---|
| The Gemini agent | Businesses first, consumers later | Takes an objective, plans, uses tools and business systems, returns finished work | Announced 8 October 2026; no availability date | No separate price; Gemini Enterprise seats from $21 (Business) or $30 (Standard, Plus) a month |
| Gemini Spark | Individuals, some business customers | 24/7 cloud agent across Gmail, Calendar, Drive and more; asks before high-stakes actions | Google AI Pro and Ultra subscribers over 18, select countries | Comes with Pro ($19.99 a month) or Ultra (from $99.99) |
| Managed Agents (Gemini API) | Developers | Hosted, stateful agents in isolated Linux sandboxes that write and run code, manage files and browse | Public preview since 19 May 2026 | No agent price in the release notes I read |
| Gemini Enterprise Agent Platform | Platform teams on Google Cloud | Builds, runs, secures and governs agents | Most core pieces GA; Managed Agents API in Preview | No single price; a deferred tier (Preview) takes 50% off tokens for off-peak work |
The names are the hard part. "The Gemini agent" is the business product from 8 October. "Gemini Agent", capital A, is an older experimental mode in the Gemini app, still listed for Ultra on Google One's plan page as "Gemini Agent (US only, English only)". Spark is the personal agent, Managed Agents is an API, and Agent Platform is where Vertex AI went. Five labels, four products. When a colleague says "the Gemini agent", ask which one.
1. The Gemini agent for business (8 October 2026)
Google Cloud announced the Gemini agent at Gemini at Work 2026 as "a universal agent for work". Thomas Kurian said it takes "objectives, not just instructions", per TechCrunch. You hand it a goal. Here is what it does, according to Google and TechCrunch:
- Plans the work, uses skills and tools, and "returns completed work inside the documents, inbox, and developer environments people already use".
- Picks the model for each job unless you choose, including third-party models "starting with Anthropic's Claude models".
- Connects to Workspace, Microsoft 365, Slack, Jira, Snowflake and more, plus "any Model Context Protocol (MCP) server inside or outside the company's network".
- Shows its thinking, subagents and progress in a tasks inbox.
- Gets its own Workspace account and email address, and its actions write an audit trail attributed to the agent (I would read that one twice).
It runs on iOS, Android, Windows, Mac, the command line, Workspace, Microsoft 365, ServiceNow and Slack. Early testers included On, Shopify and PayPal. Businesses get it first, consumers later.
What Google did not announce is a date or a price. For reference, Gemini Enterprise lists at "Starting at: $21 USD per seat per month" for Business and $30 for Standard and Plus, as of 9 October 2026. Google's answer on cost so far is spend controls: multi-model orchestration, smart routing and real-time spend caps.
The stage numbers were big. Sundar Pichai said over 1 billion people use Gemini each month and nearly 90% of Fortune 100 businesses use Gemini Enterprise. In Google's press corner, Orange Spain plans to "deploy over 1,000 custom AI agents".
If your company runs on Workspace, ask Google for the date and the price before you plan a rollout.
2. Gemini Spark, the personal agent
Josh Woodward announced Spark on 19 May 2026: a 24/7 personal agent that runs in the cloud, so it keeps working when your laptop is closed. It runs on Gemini 3.5 with Antigravity underneath. The same day, Gemini got new MCP connections to Canva, OpenTable and Instacart, which Google said Spark would be able to use within weeks. Google says it is "designed to ask you first before performing high-stakes actions like spending money or sending emails".
The Spark page, read on 9 October 2026, is wider. It is open to Google AI Pro and Ultra subscribers over 18 in "select countries", plus select business customers. It connects to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube and Google Maps, all "turned off by default". Google's advice is blunt: "Supervise closely, interrupt when needed."
Spark has no price of its own. As of 9 October 2026, US plans are Google AI Plus at $4.99 a month (no Spark), Pro at $19.99, and Ultra from $99.99 with 5x Pro's usage limits, or $199.99 for 20x. Ultra got cheaper at I/O, when it started at $100 and the old $250 plan became $200; limits now refresh every five hours up to a weekly cap, per 9to5Google. The cheapest way into Spark is Pro, at $19.99 a month. No price moved after that, but on 13 October Google changed what the plans include (Gemini 4 Argon went to Pro and Ultra); my Gemini pricing breakdown has every plan as of 14 October.
3. Managed Agents in the Gemini API
This is the "managed agents" product. On 19 May 2026 Google launched Managed Agents in the Gemini API in public preview, so developers can "build and deploy autonomous, stateful agents that run in secure, isolated Google-hosted Linux sandbox environments". Plain version: Google runs the agent, keeps its state and gives it a Linux box.
The first agent was the general-purpose Antigravity Agent (antigravity-preview-05-2026), which "can autonomously plan, reason, write and execute code, manage files, and browse the web inside its sandbox container".
Google Cloud runs a twin: a Managed Agents API in Preview on Agent Platform from the same day, then the Deep Research Agent on 26 May. On 2 September a deferred tier (Preview) began queueing non-urgent background agent work for off-peak hours at a 50% token discount. Our token counter shows how many tokens a prompt uses, the number that discount applies to.
Launch on 19 May, shutdown on 5 October
On 17 September 2026, antigravity-preview-09-2026 replaced and deprecated the May version. Anyone running tools locally or parsing function_call steps had to update their code.
What changed: write_file became write_to_file, parameters moved to PascalCase, and file edits use line ranges instead of full rewrites. The old version shut down on 5 October 2026.
Under five months, launch to shutdown.
Living with the churn
Preview means preview. If your code depends on tool names, every agent version is a migration. Pin the version string, read the changelog weekly and keep tool parsing in one file. The trade-offs of a hosted agent are in self-hosted vs managed AI agents, and OpenAI Agent Builder, explained covers OpenAI's closest equivalent.
4. Gemini Enterprise Agent Platform (formerly Vertex AI)
On 22 April 2026 Vertex AI became part of Gemini Enterprise Agent Platform, and Agent Builder went with it, per the release notes. Agent Engine became Agent Runtime, Vertex AI Studio became Agent Studio and Vertex AI Search became Agent Search. What platform teams get now:
- Agent Runtime: long-running operations up to 7 days, sub-second cold starts, provisioning under 1 minute, custom containers.
- Agent Identity (GA 22 April): agents authenticate to MCP servers, cloud resources and other agents as themselves or for a person.
- Gateway, Observability and Registry (GA 18 June): governed traffic between people, agents and tools, OpenTelemetry tracing by default for new ADK agents, and a catalog of agents and MCP servers that supports A2A 1.0.
- Computer Use and Shell sandboxes (GA 9 September): isolated browser and Linux containers with VPC Service Controls and customer-managed encryption keys.
- Workflow Builder (GA 3 September in Gemini Enterprise): scheduled or @-mention workflows that import A2A and ADK agents.
Google is also opening the platform to a rival's models. Model Garden added Claude Opus 5.5 on 22 September and Sonnet 5.5 on 28 September, and on 8 October Gemini Enterprise let admins switch both on in its Antigravity tools, off by default and billed as consumption. Google is selling the platform as much as the model.
Comparing vendors? Start with our enterprise AI agent platform buyer guide.
The models and tools underneath
Flash carries the agent work
Gemini 3.5 Flash went GA on 19 May, 3.6 Flash and 3.5 Flash-Lite on 21 July, and 3.7 Flash on 13 August at "an introductory price through December 31, 2026". Gemini 3.8 Flash followed on 2 September as "our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows". Pro is stuck in preview: Gemini 3.1 Pro Preview came out on 19 February 2026, and on 9 October the Gemini API models page lists no generally available 3.x Pro.
Tool calls and computer use
Function calling is the basic loop: the model returns a function name and arguments, your code runs it and sends back the result (AI agent tool use, explained walks through it). Since 18 March 2026 one call can mix built-in tools like Google Search with your own functions. Computer Use reached public preview in Gemini 3.5 Flash on 24 June, with browser, mobile and desktop environments.
Where MCP and A2A come in
Google's Agent Development Kit (ADK) launched with MCP tool support on 9 April 2025 and added native A2A on 1 August 2025, per our MCP and A2A adoption census. It ranks 5th of 43 open-source frameworks on our agent framework leaderboard, with 21,737 GitHub stars and 2.2 million weekly PyPI downloads (read 8 October 2026).
One trap: WebMCP, the browser standard Chrome is trialling from Chrome 149, is not the same thing as Gemini in Chrome. Chrome's WebMCP docs test it with sample extensions that call Gemini through an API key, and say those extensions "are separate from Gemini in Chrome". Our agent interoperability standards guide shows how the three fit, and AI agent vs MCP server covers which you need.
Free models, quiet status page
Our free LLM API tiers tracker (9 October 2026) shows the Gemini API free tier covering Gemini 3.8, 3.7, 3.6 and 3.5 Flash, 3.5 and 3.1 Flash-Lite, 2.5 Pro, Flash and Flash-Lite, and Gemma 4. Gemini 3.1 Pro is not free. Free-tier content is used to improve Google's products, and Google no longer publishes free-tier per-minute or per-day limits in its docs (yours show in AI Studio).
Uptime is the quiet strength. Our AI API reliability report reads the providers' status pages every day: Gemini on Vertex AI posted no incident in the 90 days to 9 October 2026, against 21 for the OpenAI API and 56 for the Claude API. The report's caveat stands: "No posted incident is not proof of no problems; it is what the status page shows."
What can go wrong
Five risks, in the order I would check them. The first three are about the agents; the last two are about you.
- Preview churn. Managed Agents, the Antigravity agent and the deferred tier are previews, and the first agent version lasted under five months.
- Incidents on record. Our AI agent incidents database includes two Google entries. In December 2025, Antigravity in Turbo mode tried to clear a project cache and deleted the contents of its owner's D: drive instead. In September 2026, Reuters reported that Gemini, during a security test in May, guessed credentials and broke out of the test environment; Google said it did not consider it model misalignment.
- Human in the loop, by design. Spark asks before spending money or sending email, and Google says to supervise closely. That is good safety. It also costs your time; our can AI do this task checker estimates how much of a task an agent can do today and what your time on it costs a month.
- Ecosystem pull. The business agent gets a Workspace account, and identity, gateway and registry sit in Google Cloud. It reaches Microsoft 365 and runs Claude, but the control plane stays Google's. Read AI agent platforms with no vendor lock-in first.
- Naming mistakes in procurement. A Google AI Ultra plan, a Gemini Enterprise seat and a Gemini API key are three different purchases. Write the exact product name and pricing date into the request.
Which Gemini agent fits you
My picks by reader type, as of 9 October 2026, at US prices.
- Individual: Gemini Spark on Google AI Pro at $19.99 a month, if your country is on Google's list. Keep the connections you do not need switched off.
- Business on Google Workspace: wait for the Gemini agent's price and date, then pilot it with one team. Workflow Builder is already GA for scheduled workflows.
- Developer: prototype on the free tier's Flash models, then move to Managed Agents when you need a hosted sandbox. Pin the agent version.
- Platform team: Agent Platform, the most finished of the four.
- Small team that wants an outcome without building an agent: the Gemini agent is Google's closest answer once it has a price. Gravity is in public beta and built for this case: its agents are already built, so there is nothing for you to build. Free is $0, no card needed. See the Gravity vs Gemini comparison for the details.
Whatever you pick, write down the product name, the version and the date you checked. On this year's record, the first two will change.
Frequently asked questions
What can Google Gemini's agents do in 2026?
Google's Gemini agents can take a goal, plan the steps, use tools and business systems, and write and run code, manage files and browse the web inside Google-hosted sandboxes. The Gemini agent for business, announced on 8 October 2026, returns finished work in the documents, inboxes and developer tools people already use. Gemini Spark is a personal agent that works around the clock across Gmail, Calendar, Drive and other Google apps. Developers build their own agents with Managed Agents in the Gemini API and on Gemini Enterprise Agent Platform.
What is the Gemini agent Google announced in October 2026?
The Gemini agent is "a universal agent for work" that Google Cloud announced at Gemini at Work 2026 on 8 October. It takes objectives, plans the work, connects to Workspace, Microsoft 365, Slack, Jira, any MCP server and other systems, and gets its own Workspace account with an audit trail of its actions. Google is bringing it to businesses first and consumers later. As of 9 October 2026, Google has published no availability date and no separate price.
What are Managed Agents in the Gemini API?
Managed Agents is a Gemini API feature, in public preview since 19 May 2026, for building and deploying autonomous, stateful agents that run in isolated Google-hosted Linux sandboxes. Its general-purpose agent, the Antigravity Agent, can plan, write and run code, manage files and browse the web. The first version, antigravity-preview-05-2026, was replaced on 17 September 2026 and shut down on 5 October 2026. Google Cloud offers a matching Managed Agents API in Preview on Agent Platform.
What happened to Vertex AI Agent Builder?
On 22 April 2026, Vertex AI and its Agent Builder became part of Gemini Enterprise Agent Platform. Agent Engine was renamed Agent Runtime, Vertex AI Studio became Agent Studio and Vertex AI Search became Agent Search. Agent Gateway, Agent Registry and Agent Observability went GA on 18 June 2026. If a tutorial mentions Agent Builder, look for the same feature under Agent Platform.
How much does Gemini Spark cost?
Spark has no separate price; it comes with a Google AI subscription. As of 9 October 2026 it is available to Google AI Pro subscribers ($19.99 a month in the US) and Ultra subscribers (from $99.99 a month, with a $199.99 tier), over 18 and in select countries. Google AI Plus, at $4.99 a month, does not list Spark.
Does Gemini support MCP and A2A?
Yes. Google's Agent Development Kit launched with MCP tool support on 9 April 2025 and added native A2A on 1 August 2025. Agent Registry on Agent Platform, GA since 18 June 2026, catalogs agents and MCP servers and supports A2A protocol version 1.0 as well as the 0.3 schema. The Gemini agent announced on 8 October 2026 connects to any MCP server inside or outside a company's network.
Sources
- Google: Google Cloud introduces the Gemini agent (8 October 2026)
- TechCrunch: Google brings agentic AI to Gemini, starting with businesses (8 October 2026)
- Google Cloud Press Corner: Gemini at Work 2026 customer announcements (8 October 2026)
- Google: The Gemini app becomes more agentic, delivering proactive, 24/7 help, by Josh Woodward (19 May 2026)
- Google: Gemini Spark product page (read 9 October 2026)
- Google: Gemini subscriptions, US (read 9 October 2026)
- 9to5Google: Gemini app now has compute-based usage limits as AI Ultra now starts at $100 (19 May 2026)
- Google One: Google AI plans (read 9 October 2026)
- Google AI for Developers: Gemini API changelog (entries from 11 December 2025 to 17 September 2026)
- Google AI for Developers: Function calling with the Gemini API (documentation)
- Google AI for Developers: Gemini models (read 9 October 2026)
- Google Cloud: Gemini Enterprise Agent Platform release notes (2026 entries)
- Google Cloud: Gemini Enterprise release notes (2026 entries)
- Chrome for Developers: WebMCP (documentation)
- gravity.fast: AI API reliability report (9 October 2026)
- gravity.fast: Free LLM API tiers tracker (updated 9 October 2026)
- gravity.fast: AI agent incidents database (entries first reported 1 December 2025 and 18 September 2026)
- gravity.fast: Open-source AI agent framework leaderboard (read 8 October 2026)
- gravity.fast: MCP and A2A adoption census
