Microsoft this week delivered on its promise to make Edge an AI-powered assistant that does web chores for you. The new Copilot Actions and Journeys, now rolling out, can unsubscribe you from newsletters, book restaurant tables, and automatically group past browsing into resumable projects. The launch came just as OpenAI unveiled its own ChatGPT-powered Atlas browser, which offers a similar agent mode and the ability to recall your browsing context across sessions. For the first time, two major browsers are shipping features that don’t just display the web—they act on it.

What Just Landed in Edge and Atlas

Copilot Actions: Your New Digital Errand Runner

Actions is the headline feature in Edge’s updated Copilot Mode. Once you opt in and grant the necessary permissions, Copilot can navigate multi-step web tasks on your behalf. In practice, that means it can:

  • Scour your email inbox, identify newsletters, and click through the unsubscribe flow for each one.
  • Fill out and submit reservation or booking forms, such as making a restaurant reservation or scheduling an appointment.
  • Compare products across open tabs and draft a summary or recommendation.
  • Synthesize content from multiple pages into a single answer.

Microsoft built in a final confirmation step before any action that is irreversible—like completing a payment or purchase. For most other tasks, Copilot will show progress as it works, but it may interact with logged-in sessions to complete forms or send requests. The feature is opt-in and requires explicit permission to access browsing history, cookies, and in some cases, active credentials.

Journeys: Remembering What You Were Doing

Alongside Actions, Microsoft rolled out Journeys, which automatically groups your past browsing into topic-based projects. If you were researching a vacation, comparing laptops, or planning a dinner party across dozens of tabs and multiple days, Journeys will surface that cluster as a single card with a summary and suggested next steps. Click it, and you’re right back where you left off.

Journeys also requires your permission to scan and organize browsing history. It’s meant to solve the “tab overload” problem and reduce time spent hunting for pages you viewed earlier. You can view, archive, or delete any journey at any time.

Atlas: ChatGPT’s Browser With a Photographic Memory

OpenAI’s counterpunch is Atlas, a standalone browser with a built-in ChatGPT sidebar that goes beyond chat. Its agent mode can open tabs, click through web pages, and attempt multi-step operations while maintaining context from the conversation. Browser memories let ChatGPT recall snippets of your past browsing so you can say, “Reopen the shoes I looked at yesterday” or “Show me that blue hoodie I was using last week,” and it will pull up the relevant tabs.

Atlas is designed from the ground up as an AI-first browser. The ChatGPT sidebar is always present, and the browser can act across sites while keeping your chat context. Like Edge’s features, everything is opt-in, with toggles to disable memory and visual indicators when the assistant is active.

Your Browser as a Digital Butler: The Promise and Peril

For the Everyday User

The immediate upside is convenience. Repetitive, low-stakes chores that usually eat up small chunks of time can be handed off. Unsubscribing from a dozen newsletters, comparing flight prices, or pulling together research notes from multiple pages becomes a one-sentence command. If the agent is reliable, it’s a genuine productivity boost.

But reliability varies. Early hands-on reports show successful unsubscribes and bookings, but also incomplete tasks, incorrect form submissions, and actions reported as completed when they weren’t. As with any nascent automation, start with tasks where a mistake is annoying but not catastrophic. Save financial transactions or legally binding submissions for manual control until the technology matures.

Privacy-conscious users must weigh convenience against data exposure. Both Edge and Atlas require access to browsing history and, for agentic actions, often need session cookies or saved passwords. If you use a shared device, those “memories” could surface to the next person who logs in. The safest approach: keep agentic features off on any computer that isn’t strictly personal, and audit the permission prompts carefully—revoke access when you no longer want the assistant snooping around.

For Power Users and Researchers

If you regularly juggle complex research projects, the synthesis and summarization features are the safest and most immediately useful. Copilot Actions can already pull content from open tabs and generate a coherent summary, which beats manually compiling notes. Atlas’s browser memories similarly reduce time spent re-finding pages.

Agent mode for power users should be treated as a beta capability. Test it with rule-based, high-friction chores where outcomes are easy to verify. For instance, letting the browser log into a forum and update a status might work; trusting it to configure a router’s admin panel or file a tax form is asking for trouble. Use a password manager and two-factor authentication as a safety net—agentic flows that depend on session hijacking are less effective when they can’t bypass hardware tokens or biometric checks.

For IT Administrators

This is where the conversation turns serious. Agentic browsing inserts a new, partially autonomous layer between an employee and every web service they use. The governance questions are immediate:

  • Does an automated agent’s access to corporate intranets, SaaS apps, or cloud consoles violate existing security policies?
  • If an employee’s browser automatically fills out and submits a form using cached corporate credentials, who is accountable for errors or data leaks?
  • How do you audit what the assistant did, and can those logs feed into your SIEM?

Microsoft has not yet published full technical documentation on how Copilot Actions handles enterprise credential isolation, data residency, or retention of browsing memories. Until it does, IT teams should assume the worst: that enabled agentic features may mix personal and corporate browsing contexts, store sensitive session data in the cloud, and lack robust audit trails. The prudent move is to disable these features by default in managed environments and run a tightly controlled pilot with a small cohort, monitoring for misbehavior and edge-case failures before any broader deployment.

The Steady March Toward an AI-Powered Web

Browsers have been creeping toward this moment for years. Extensions and scripts could automate form filling and clicks, but they required user configuration and lacked natural-language interfaces. Google’s Duplex demonstrated automated phone calls, but it never materialized as a browser-level feature. The arrival of large language models changed the equation: when a browser can understand a plain-English request and translate it into a sequence of web actions, automation becomes accessible to anyone.

Microsoft fired the first shot in July with Copilot Mode in Edge, which initially offered sidebar chat and page summarization. The Actions and Journeys features were announced at the same time but only started rolling out now. OpenAI’s Atlas, launched just days ago, shows how quickly the competitive pressure is building. Both companies see the browser as the operating system for the web, and whoever controls the AI assistance layer gets to shape how billions of people interact with online services.

Before You Hand Over the Keys, Read This

If you’re an everyday user:
- Start with the read-only capabilities—summarization, chat, and Journeys—before enabling Actions.
- Keep agentic mode switched off on shared or public computers.
- Scrutinize every permission prompt; revoke access to browsing history and cookies when you’re done.
- Use a reputable password manager and hardware-based two-factor authentication wherever possible. This limits the damage if an automated session is hijacked.
- Regularly clear your browsing history or manage your “memories” in Atlas/Edge settings to keep a minimal footprint.

If you’re an IT admin or manage enterprise endpoints:
- Confirm the default state of agentic features in your fleet. They should be disabled until you explicitly enable them.
- Demand clear answers from vendors: Where are browser memories stored? For how long? Is the data used for model training, even in aggregate?
- Before any pilot, verify that the browser can isolate personal and corporate credentials and that you can export agent-action logs to your security tools.
- Begin any pilot with low-risk, high-ROI tasks like internal knowledge base searches or marketing research. Financial, legal, or HR processes should remain off-limits.
- Train your staff to recognize when the browser is acting autonomously, how to confirm or cancel actions, and how to revoke permissions if they suspect a problem.

The Browser War Has a New Battlefield

This is no longer about rendering speed or extension catalogs. The browsers that succeed will be the ones that build the most trusted, capable context layer—a personal assistant that knows what you’ve done and can act on your behalf across the web. Google is almost certainly working on similar features for Chrome, and Mozilla is exploring local-only AI models to offer privacy-preserving alternatives.

Standards bodies will need to catch up. Websites must have clear mechanisms to declare whether they allow automated agents and under what conditions. Users need a universal control panel to see and revoke permissions granted to any browser assistant. And regulators will inevitably scrutinize how these agents handle personal data, especially when they start making purchases or submitting forms with legal weight.

For now, the best advice is to embrace the productivity gains cautiously. Try the read-only features, establish personal boundaries around what you’ll let the browser do automatically, and keep a close eye on the permission toggles. The AI browser revolution is real, but for the next year or two, it will be a careful dance between automation and oversight.