What Is an AI Agent? How They Work and What They Can Actually Do
An AI agent is software that uses a language model to plan, call tools and act over many steps toward a goal. Here is how agents work and what the evidence shows they can do.
// ai agents
AI Agent Launch Tracker
A dated, sourced timeline of major AI agent launches since October 2024, from OpenAI, Anthropic, Google and others. It covers browser and computer-use agents, coding agents, agent platforms and the MCP and A2A protocols.
Researched and fact-checked by AI, with no human review. How we verify
25 entries from 26 sources. Checked for news daily; last checked
What this tracker follows
This tracker lists major AI agent launches and milestones since October 2024. It is current as of October 6, 2026. An AI agent here means software that uses an AI model to plan and carry out multi-step tasks with tools. It does more than answer questions.
Entries cover:
- Computer-use and browser agents, which operate a screen by clicking and typing.
- Coding agents.
- Platforms and software development kits (SDKs) for building agents.
- Shared protocols such as the Model Context Protocol and Agent2Agent.
The Model Context Protocol connects AI assistants to tools and data. Agent2Agent is designed to let agents from different vendors work together.
Each launch or milestone gets one entry, dated to its announcement. Each is cited to the organization behind it or to an established news outlet. The one scheduled entry is instead dated to a future shutdown date set by the organization named. Status words such as research preview, beta and generally available follow the source. Statements about what a product can do are the companies' own claims.
The list is capped at 25 entries, so the selection is editorial. Model releases, funding rounds, acquisitions, startup products and most follow-up updates are left out. That includes the other agent tools OpenAI listed in its DevDay 2026 recap of September 29, 2026. Products change after launch, so some entries note a later rename, merger or shutdown recorded on the source page.
Scheduled next
Targets set by the organizations named. Dates like these often move.
Scheduled
OpenAI to wind down Agent Builder and its Evals product
In an update dated June 3, 2026, OpenAI said it is winding down Agent Builder and Evals, which will no longer be available on its platform from November 30, 2026. Agent Builder launched in beta with AgentKit in October 2025. OpenAI points users to its Agents SDK or to Workspace Agents in ChatGPT.
Source: Introducing AgentKit, OpenAI
Timeline
OpenAI introduces dots, always-on agents in ChatGPT
OpenAI introduced dots, which it describes as always-on agents. It says they run on its GPT-6 Astra model, have their own cloud computer and browser, and can connect to over 4,000 apps. Rollout began on Pro and Business Premium plans in eligible markets, with an admin-enabled beta for Enterprise.
Source: Introducing dots, OpenAI
Meta launches Muse, a personal AI agent, in the United States
Meta introduced Muse, a personal agent it says can send email, book travel and fill out web forms. Meta says Muse runs in a dedicated cloud virtual machine, where a separate Sentinel agent must approve anything sent to the internet. Meta said it was rolling out in the US on iOS, Android and muse.ai.
Source: Introducing Muse: The World’s First Personal AI Agent Built for Everyone, Meta
MCP 2026-07-28 specification released with a stateless core
The Model Context Protocol's maintainers released the 2026-07-28 specification, which drops the opening handshake and sessions so that each request stands alone. It also adds a formal extensions framework, tighter authorization rules and a deprecation policy with a minimum twelve-month window.
Source: The 2026-07-28 Specification, Model Context Protocol project
Microsoft introduces Scout, its first always-on Autopilot agent
Microsoft introduced Autopilots, which it calls always-on agents that act under their own identity, and Scout, the first of them, which it says runs on open-source OpenClaw technology. Scout opened in private preview and as an experimental Frontier release. A later note says it is now Autopilot.
Source: Introducing Microsoft Scout: Your always-on personal agent, Microsoft
Google introduces Gemini Spark, a personal agent in the Gemini app
For I/O 2026, Google introduced Gemini Spark, which it calls a 24/7 personal agent. Google says the cloud-based agent keeps working when a laptop is closed or a phone is locked and connects to Gmail, Docs and Slides. It planned access for trusted testers that week and a U.S. AI Ultra beta the next.
Source: The Gemini app becomes more agentic, delivering proactive, 24/7 help, Google
A2A project reports 150+ backers and a stable 1.0 specification
At its one-year mark, the Linux Foundation-hosted A2A project said more than 150 organizations support the protocol and that version 1.0, its first stable specification, was out. It cited integrations in Microsoft's Azure AI Foundry and Copilot Studio and in Amazon Bedrock AgentCore Runtime.
Anthropic opens Claude Managed Agents in public beta
Anthropic launched Claude Managed Agents, a set of APIs for building and running cloud-hosted agents on infrastructure it manages, in public beta on the Claude Platform. Pricing was standard token rates plus $0.08 per session-hour of active runtime. Multi-agent coordination was a research preview.
Source: Claude Managed Agents: get to production 10x faster, Anthropic
OpenAI introduces Frontier, an enterprise platform for AI agents
OpenAI introduced Frontier, a platform for enterprises to build, deploy and manage AI agents with shared business context and set permissions. It was available to a limited set of customers at launch. OpenAI named HP, Intuit, Oracle, State Farm, Thermo Fisher and Uber among the first adopters.
Anthropic launches Cowork as a research preview
Anthropic announced Cowork, a tool in the Claude desktop app. It lets Claude read and change files in a folder the user chooses, bringing the approach of Claude Code to non-coding work. TechCrunch reported that it launched as a research preview for Max subscribers, with a waitlist for other plans.
Source: Anthropic’s new Cowork tool offers Claude Code without the code, TechCrunch
Linux Foundation forms Agentic AI Foundation; MCP is contributed to it
The Linux Foundation announced the Agentic AI Foundation with three founding projects: the Model Context Protocol from Anthropic, the goose agent framework from Block and AGENTS.md from OpenAI. Platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
AWS announces three frontier agents, including Kiro autonomous agent
AWS announced three of what it calls frontier agents for software teams: Kiro autonomous agent for development work, AWS Security Agent and AWS DevOps Agent. AWS says agents in this class can work for hours or days without intervention. All three were made available in preview the same day.
Microsoft announces Agent 365 to manage and secure AI agents
On the first day of Ignite 2025, Microsoft announced Agent 365, which it calls a control plane for AI agents: a registry, access controls, dashboards and security tools. It pointed IT teams to its Frontier early-access program. A note since added to the post says it is generally available.
Source: Microsoft Agent 365: The control plane for AI agents, Microsoft
Amazon Bedrock AgentCore becomes generally available
AWS made Amazon Bedrock AgentCore, its platform for building, deploying and running AI agents on any framework or model, generally available in nine AWS Regions. The release added Agent2Agent protocol support to AgentCore Runtime and let AgentCore Gateway connect to existing MCP servers.
Source: Amazon Bedrock AgentCore is now generally available, Amazon Web Services
Google releases Gemini 2.5 Computer Use model in public preview
Google released a specialized model built on Gemini 2.5 Pro that lets agents click, type and scroll through user interfaces. It opened in public preview through the Gemini API in Google AI Studio and Vertex AI. Google said it is tuned mainly for web browsers, not yet for desktop operating systems.
Source: Introducing the Gemini 2.5 Computer Use model, Google
OpenAI launches ChatGPT agent, merging Operator and deep research
OpenAI launched ChatGPT agent, which it said uses its own virtual computer to browse websites, run code and produce slides and spreadsheets, combining Operator's web actions with deep research. Pro, Plus and Team users could switch on agent mode. OpenAI has since marked the launch post as outdated.
Source: Introducing ChatGPT agent: bridging research and action, OpenAI
Microsoft announces GitHub Copilot coding agent and MCP support at Build
At Build 2025, Microsoft announced an asynchronous coding agent for GitHub Copilot. It also announced Model Context Protocol support across GitHub, Copilot Studio, Dynamics 365, Azure AI Foundry, Semantic Kernel and Windows 11. It said Microsoft and GitHub had joined the MCP Steering Committee.
Source: Microsoft Build 2025: The age of AI agents and building the open agentic web, Microsoft
OpenAI launches Codex, a cloud coding agent, as a research preview
OpenAI launched a research preview of Codex, a cloud-based software engineering agent that works on many tasks in parallel, each in its own sandbox loaded with the user's code repository. It ran on codex-1, a version of the o3 model. Rollout began with ChatGPT Pro, Enterprise and Business users.
Source: Introducing Codex, OpenAI
Google announces the Agent2Agent (A2A) protocol
Google announced Agent2Agent, an open protocol to let AI agents built by different vendors exchange information and coordinate work, with support from more than 50 technology partners. Google presented it as a complement to Anthropic's Model Context Protocol and published a draft specification.
Amazon introduces Nova Act and a research preview of its SDK
Amazon introduced Nova Act, an AI model it said was trained to carry out actions in a web browser. It also released a research preview of the Nova Act SDK, so developers could build browser agents on an early version of the model. U.S. customers could get the SDK through the new nova.amazon.com site.
Anthropic introduces Claude Code in a limited research preview
Alongside its Claude 3.7 Sonnet model, Anthropic introduced Claude Code, a command line tool for agentic coding. Anthropic said it can read and edit code, write and run tests, and commit and push changes to GitHub from a developer's terminal. It was released as a limited research preview.
OpenAI releases Operator, a browser agent, as a research preview
OpenAI released Operator, an agent that uses its own browser to type, click and scroll through websites, to ChatGPT Pro users in the U.S. as a research preview. It ran on a model called Computer-Using Agent. A July 17, 2025 update on the page says Operator was merged into ChatGPT as ChatGPT agent.
Source: Introducing Operator, OpenAI
Google introduces Project Mariner and Jules alongside Gemini 2.0
With Gemini 2.0, Google presented agent prototypes. One was Project Mariner, an early research prototype that completes tasks in the browser through an experimental Chrome extension. Another was Jules, an experimental coding agent that works inside GitHub. Trusted testers were the first to try Mariner.
Source: Introducing Gemini 2.0: our new AI model for the agentic era, Google
Anthropic open-sources the Model Context Protocol (MCP)
Anthropic open-sourced the Model Context Protocol, an open standard for linking AI assistants to data sources and tools. The release included a specification and SDKs, local MCP server support in the Claude desktop apps and pre-built servers for systems such as Google Drive, Slack and GitHub.
Anthropic releases computer use in public beta
Anthropic released computer use in public beta with an upgraded Claude 3.5 Sonnet, letting developers have Claude look at a screen, move a cursor, click and type. It was offered on Anthropic's API, Amazon Bedrock and Google Cloud Vertex AI. Anthropic called it experimental and at times error-prone.
Source: Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku, Anthropic
Additional sources
- DevDay 2026 Recap, OpenAI
Update history
- Rewritten in shorter, plainer sentences. No facts were changed.
Articles on AI Agents
Agentic AI Reality Check: Where Agents Work Today and Where They Fail
Agentic AI is furthest along in software development at large companies. Elsewhere most projects remain pilots, reported profit impact is flat and security controls trail autonomy.