Skip to main content

Agents in the Apps You Already Pay For

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.

In 2025 the advice was to pick an agent product: Operator, Atlas, Mariner, Comet. By October 2026 most of those are gone, and what replaced them is a mode inside the assistant you already subscribe to. Claude folded its Cowork agent into the main app. ChatGPT shut down its Atlas browser and shipped ChatGPT Work and Dots. Gemini absorbed Project Mariner into Agent Mode. Microsoft rebuilt Copilot around agents. Meta launched Muse. All of them are bundled into the roughly twenty-dollar tiers rather than sold separately.

This page is the beginner on-ramp to the track. You will learn what the agent mode in your app can do, what it still fails at, and the three controls (approvals, scoping, blast radius) that matter more than any product's UI.

Learning focus​

  • Recognize the three shapes of consumer agent: works-in-my-files, browser, and always-on
  • Know what the agent mode in each major app does as of October 2026
  • Set approvals, folder and connector scope, and blast radius before the first run
  • Spot the tasks these agents are still bad at

The three shapes​

Every consumer agent in 2026 is one of three shapes, and most apps now offer all three.

The works-in-my-files-and-apps agent takes a goal, reads your documents and connected services, and produces something: a spreadsheet, a deck, a cleaned-up folder, a batch of drafted replies. This is what Claude (formerly Cowork) and ChatGPT Work do.

The browser agent drives a web browser to research, compare, fill in forms, and click through sites that have no API. Claude in Chrome, ChatGPT's Chrome extension, and Gemini's auto-browse in Chrome are the survivors; the standalone AI browsers (Atlas, Mariner) were retired in 2026.

The always-on agent runs in the cloud on a schedule or a trigger while your laptop is closed: Claude's scheduled tasks and Routines, ChatGPT's scheduled tasks and Dots, Gemini Spark. This is the newest shape and the one most likely to generate "how do I set up a recurring agent" questions in 2027.

What each app offers​

Treat the product names below as a snapshot; the shapes above are the durable part. Where a fact rests on press reports rather than the vendor's own pages, it is marked as reported.

Claude​

Anthropic's pricing page now carries the line "Claude Cowork is now just Claude," rolling out to Pro and Max. The agent takes a goal and works across your files and tools, runs multi-step tasks in parallel, shows each step so you can stop or redirect, and can hand a task off to run unattended in the cloud, including from your phone. You choose which folders it can see and can mount them read-only or read-write-without-delete; any deletion needs your approval. It uses sub-agents, skills, and plugins, and Anthropic's own page notes it consumes usage limits faster than chat.

Claude in Chrome is available on all paid plans; on Max and Team it runs as a full agent session saved to your history. The side panel can run in an automatic-approval mode with pauses for consequential steps, and Anthropic's help article says the product is "enhanced with our safety classifiers but is still risky." Financial-services sites are blocked by default, and the guidance is to start with trusted sites and avoid sensitive financial, legal, or medical ones. Claude Tag, an "@Claude" teammate in Slack channels, is in beta for Team and Enterprise. Memory and connectors are available on the free tier; the agent features start at Pro. Anthropic merged chat and Cowork memory in late August 2026 (reported).

ChatGPT​

ChatGPT Work launched on July 9, 2026 as the agent mode on paid plans other than Free and Go. Reviewers describe it as a background worker: it takes a brief, works for minutes or hours across your connected apps, and hands back a finished file (reported). Scheduled tasks were rebuilt in June 2026 and run at most once per hour. At DevDay on September 29, 2026 OpenAI added Dots, persistent agents that keep working on ongoing tasks and check in through chat tools, alongside ChatGPT Space, a shared workspace for teams and agents; Dots are reported to be included with Pro tiers and Business Premium rather than Plus (single source, unverified).

Two retirements matter for anyone following older guides. The Atlas browser stopped working on August 9, 2026, with browsing moved into the redesigned desktop app and a Chrome extension (reported). And since August 2026 individual accounts can no longer create Custom GPTs; OpenAI's replacement is Workspace Agents on Business, Enterprise, and Edu plans.

Gemini​

Google lists Agent Mode as an early-access feature on paid Gemini tiers. Project Mariner, Google's standalone browsing agent, was shut down on May 4, 2026 and folded into Gemini and Chrome's auto-browse (reported). At I/O in May, Google announced Gemini Spark, a personal agent that runs continuously on dedicated cloud machines, shipping first to US AI Ultra subscribers (reported). In October, Google Cloud introduced a Gemini agent for work in enterprise preview, with "coworker" agents that get their own Workspace account and an audit trail attributed to the agent rather than a person (reported, not confirmed from Google's own pages). Gemini's Deep Research can read your Gmail, Drive, and Chat with permission. Note that since October 9, 2026 the free Gemini tier is limited to Flash-Lite (reported), so the agent features effectively start at a paid plan.

Microsoft Copilot​

Microsoft relaunched Copilot as a "super app" on September 25, 2026 with Home, Code, and Autopilot tabs (reported, not confirmed from Microsoft's own pages). For consumers, Copilot Pro was discontinued in August 2026 and its features moved into Microsoft 365 Premium (reported). For organizations, Microsoft 365 Copilot remains a per-user add-on, and Microsoft Agent 365, generally available since May 1, 2026, provides identity, inventory, and controls for agents at a per-user price, with a free foundational plan for Microsoft Cloud customers. One analyst estimate puts total runtime cost at two to three and a half times the headline seat price once compute is billed (secondary estimate, not Microsoft's figure).

Meta Muse​

Meta announced the Muse personal-agent app on September 8, 2026 with free, Power, and Maximum tiers, US only at launch, on iOS, Android, the web, and WhatsApp (reported; Meta confirmed prices to CNBC). On October 12 Meta released the Muse Wonder models behind it, marketed on "Active Memory" and "Active Focus," the latter described as not giving up when it hits a blocker. Meta had earlier acquired Manus, a general-purpose task agent, in late 2025 (reported).

The three controls that matter​

Vendors converged on the same safety posture in 2026, and it is more useful to learn than any one app's menus.

Approvals. Decide before the run which actions need a yes: sending messages, purchases, deletions, anything that leaves your account. Every app above offers an approval mode; the trap is approval fatigue. Anthropic reported that users approved about 93 percent of permission prompts in Claude Code, which is why its auto mode moved to a classifier for routine actions. Fewer, better-placed approvals beat many reflexive ones. Ask for approval only at irreversible steps and read those ones properly.

Scoping. Give the agent the smallest view that fits the task. Pick a folder, not your whole drive. Grant the one connector the job needs, not all of them. Prefer read-only access where the task is research. Claude's folder mounts and Routines' per-connector lists exist for exactly this; Anthropic's Routines documentation warns that an included connector lets the agent use every tool it exposes, including writes, without asking.

Blast radius. Ask what the agent could change or delete if it misreads the goal, and shrink that before you start. Keep agents out of financial, legal, and medical sites and apps, as every vendor advises. Close sensitive apps before a computer-use session. Treat anything the agent reads (a web page, an email, a calendar invite) as potentially hostile, because 2026 security research showed zero-click hijacks of browser agents through exactly those channels. Agent Safety and Guardrails walks through the incidents.

Your first delegation​

Pick a task you do weekly that touches files but not money. Write a brief with six parts and run it attended the first time.

Objective: [the outcome, in one sentence]
Context: [who it is for, why it matters, what good looks like]
Resources: [the folder, files, or connectors it may use; nothing else]
Constraints: [read-only unless stated; ask before sending or deleting; stop after 20 minutes]
Output: [the file or message you want, in what format, saved where]
Verification: [how you will check it: spot-check three items, compare to last week]

Watch the steps as it works. Where it guesses, add a line to the brief. After two failed corrections, stop and rewrite the brief rather than arguing with the run. Once it works attended, schedule it and spot-check the first three runs.

What they still fail at​

Reviewers of both Claude's and ChatGPT's agents report the same breaking points: website logins, long unattended runs, and anything that must be repeatable and auditable (reported). Anthropic says its agent is not designed for checkout tasks. Vendors themselves describe browser and computer-use agents as risky. If a job must happen the same way every time with a record of each step, a workflow tool or a scheduled script is the better fit; Workflows vs Agents covers that decision.

Key takeaways​

  • The agent is a mode inside the assistant you already pay for; standalone agent products mostly disappeared in 2026.
  • Three shapes recur everywhere: works-in-my-files, browser, and always-on.
  • Learn approvals, scoping, and blast radius; they transfer across every app and outlast product names.
  • Agents still break on logins, long unattended runs, and anything that must be auditable. Keep those tasks closer.

Next steps​

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks