AI Frontier

Always-on agents

The next product after the chatbot: an agent that runs on its own computer, works toward a goal you set, and keeps going after you close the app. Verified November 4, 2026.

1 of 4

Goals, not prompts

You describe an outcome and the limits on what the agent may do by itself. It plans the steps and keeps going without a new message each time.

2 of 4

Its own computer

Work runs on a cloud machine or server the agent controls, so it continues after your laptop is closed. Each vendor isolates this differently.

3 of 4

Triggers and schedules

Agents can wake on a timer, an incoming email, a code change, or an alert, rather than waiting for you to open an app.

4 of 4

Approval gates

Every vendor-run product here pauses for sign-off on consequential actions — purchases, sending on your behalf, password changes. Where that line sits is the main design choice.

At a glance

Four of the five big US labs shipped an always-on agent between May and September 2026, and they converged on the same design: a cloud computer per user, a goal instead of a prompt, and human approval before anything consequential. They differ on who pays (Meta's is free; OpenAI's starts at $100 a month), on where they reach you (WhatsApp vs Slack vs Gmail), and on how the agent is kept from misusing your credentials. None has a shared independent benchmark, so comparisons are about access, price and controls rather than capability.

Agent by agent

Dots

OpenAI · Sep 29, 2026 (DevDay)

Limited rollout

You give a dot a goal and limits on what it may do alone; it keeps working in the background and asks when it needs approval or more detail. Reaches more than 4,000 apps through OpenAI's plugins.

Model
GPT-6 Astra
Runs on
A cloud computer and browser per dot
Where you reach it
ChatGPT, Slack, Microsoft Teams, voice calls; text messages planned
Who can use it
First dot included with ChatGPT Pro ($100, $200, $500/mo) or a Business Premium seat ($125/mo, $100 billed annually). Enterprise and Edu in admin-enabled beta. Pro rollout excludes the EEA, Switzerland and the UK.
Controls
Signs in with saved passwords the model never sees. Sensitive actions such as changing a password stay with the user. Specialist dots for companies get their own identity and credentials, with Microsoft Agent 365 governance in preview.

Worth noticing

When idle, a dot does what OpenAI calls proactive research. OpenAI also says it wants teams of dots working together — today each plan includes one.

YourStory launch coverageImplicator: plans and pricing

Muse

Meta · Sep 8, 2026

Available

A personal agent for everyday tasks — sending email, booking travel, shopping, planning — that builds an action plan from a goal and keeps working after the app is closed.

Model
Muse Spark (Muse Wonder to follow)
Runs on
Muse Secure VM — a dedicated cloud virtual machine per user
Where you reach it
iOS and Android apps, the web, WhatsApp; AI glasses later
Who can use it
US, ages 18+. Free tier, plus Power ($20/mo) and Maximum ($100/mo) plans with higher limits.
Controls
A separate Sentinel agent must approve anything Muse sends to the internet. Credentials are stored so Muse can use them without seeing passwords or payment details. Users choose connected apps and can review an activity log.

Worth noticing

The only free always-on agent from a major lab. Meta is also exploring taking a share of purchases the agent makes, which makes the shopping use case a business model as well as a feature.

VKTR launch coverageRiviera Radio: security and monetization

Grok Bot

SpaceXAI (xAI) + Cursor · Aug 11, 2026

Beta

A team of always-on AI teammates that sign into your work tools — including sites with no API — finish tasks end to end, and check back only when they need approval. Pitched at sales, support, recruiting, engineering and ops.

Model
Grok models
Runs on
A cloud computer per bot, on Cursor's infrastructure
Where you reach it
Mac and iPhone apps first; Windows and Linux available; Android to follow
Who can use it
SuperGrok Heavy, Cursor Ultra, or Cursor Teams Premium subscribers — plans from roughly $120 to $300/mo.
Controls
Approval checkpoints before consequential actions. xAI's documentation tells users to review approvals, security and privacy settings before connecting sensitive systems, and has not detailed how sign-in credentials are handled.

Worth noticing

The first joint product of xAI and Cursor. Developed inside Cursor under the codename 'Sand', it ships through Cursor's checkout while SpaceX's $60B purchase of Cursor's parent, Anysphere, heads toward closing.

DataCamp explaineriGeeksBlog: platforms and pricing

Gemini Spark

Google · May 19, 2026 (I/O)

Limited rollout

A 24/7 personal agent that works across Gmail, Calendar, Docs, Sheets, Slides and Drive even when your phone and laptop are off. Third-party apps connect over MCP, starting with Canva, OpenTable and Instacart.

Model
Gemini 3.5 Flash on the Antigravity agent harness
Runs on
Dedicated Google Cloud virtual machines
Where you reach it
Gemini app, Android (live status in Android Halo); Chrome planned
Who can use it
Google AI Pro and Ultra subscribers in the US.
Controls
Checks with the user before major actions. Purchases need explicit approval at launch; Google's Agent Payments Protocol (AP2) adds spending limits and approval rules for later autonomous payments.

Worth noticing

The always-on agent with the deepest first-party data access: it starts inside the user's Google Workspace rather than having to be connected to it.

Digital Applied launch analysisSiliconANGLE

Cursor Automations

Cursor (Anysphere) · Mar 5, 2026

Available

Coding agents that start themselves. A trigger — a code change, a Slack message, a timer, or a PagerDuty incident — launches an agent that reviews code, investigates issues, or does recurring engineering chores.

Model
Your choice of models in Cursor
Runs on
An isolated cloud sandbox spun up per trigger
Where you reach it
Cursor editor, Slack, iOS app with Live Activities
Who can use it
Cursor subscribers.
Controls
Each run is sandboxed with access only to the codebase and services you connect over MCP (Datadog, Linear, Notion, Slack). A central view shows every agent's actions.

Worth noticing

The developer version of the same idea, ahead of every consumer launch on this page. Cursor says it runs hundreds of automations an hour on its own codebase.

ChatAI summaryMachine Herald

OpenClaw

Open-source project · Late 2025

Available

The open-source agent that started the category: a self-hosted runtime that stays on, watches your channels, runs scheduled and triggered tasks, and is extended with community skills.

Model
Any — GPT, Claude, Gemini, or open-weight models
Runs on
Your own computer or server (local machine, VPS, or managed host)
Where you reach it
WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Teams and more
Who can use it
Free, MIT licence. You pay for the model API and the machine.
Controls
Whatever you configure. The agent has the same file and network access as the machine it runs on, which security teams flag as a risk on work laptops.

Worth noticing

The reference point for every product on this page: Fox Business described Meta's Muse as modeled on OpenClaw, and Gizmodo framed Gemini Spark as Google coming for it.

OpenClaw on GitHubAona: enterprise security view

Conway

Anthropic · Not announced

Not launched

An always-on Claude agent that surfaced in leaked Claude Code source in March and was tested internally, reportedly ending on July 24. Reports describe scheduled and webhook triggers, Chrome control, and Claude Code inside one persistent environment.

Model
Claude
Runs on
Reported: a standalone always-on Claude environment
Where you reach it
Reported: its own interface, browser control, webhooks, notifications
Who can use it
Internal test only. Anthropic has not announced Conway or a release date.
Controls
Not published.

Worth noticing

Anthropic's shipping pieces are developer building blocks rather than a consumer agent: Claude Managed Agents (hosted agent runtime, $0.08 per session-hour plus tokens, since April) and Claude Code Routines.

TestingCatalog: test endingLeak analysis

Prices are monthly subscription prices in USD as reported at launch; plans and regions change often, so check the linked sources before relying on them. Capability and safety claims are the vendors' own — there is no shared independent benchmark for always-on agents yet.