Browsing

Web Browsing — Agents That Read the Live Web

A real browser, not an HTTP fetch. Agents open JavaScript-rendered pages, extract readable content, drive multi-step sessions, and hand back artifacts as links — with the credentials resolved outside their reach.

Renders the page, not the HTML

A plain HTTP GET returns an empty shell on most modern sites. Forge drives a real Chrome session, so an agent reads what a person would see — SPAs, dynamic listings, booking and travel sites included.

Readable extraction, not raw dump

Pages come back readability-cleaned rather than as whole-body text, so an agent spends its context on the article instead of the navigation.

Multi-step sessions

Beyond single-shot fetching: open a session, act across several steps, close it. Screenshots and files produced along the way are stored and returned as links.

The key never reaches the agent

Browsing runs through a credential-fronted service. The agent asks for a page; it never holds the key that made the request possible.

Why a browser and not a fetch

Most of the web an agent needs is no longer in the HTML you get from a GET request. Listings, dashboards, booking flows and half the modern press render client-side. An agent limited to raw HTTP concludes the page is empty — and, worse, often summarises that emptiness confidently.

So Forge browses with an actual browser. The agent asks for a URL and receives what a person sitting in front of the page would see.

What agents can do with it

  • Read a rendered page — the common case, and the one that makes research trustworthy. This is what lets a verifying agent open a source and confirm a quote rather than trusting a link’s title.
  • Work across several steps — open a session, navigate, act, collect. Useful when the thing you want is three clicks deep rather than at a URL.
  • Bring back artifacts — screenshots and files produced during a session are stored and handed back as links, so a large capture never has to be pasted into a transcript.

Where it earns its place

Browsing is not a feature people ask for by name. It shows up as the reason other things work:

  • The research use case verifies sources by opening them. Without a real browser, “verified” would mean “the URL responded.”
  • A digest workflow reads sources that render client-side, which is most of them.
  • A planning workflow can read live prices and schedules rather than a model’s recollection of them.

Boundaries worth knowing

  • Egress is watched and scoped. Browsing is an outbound capability, and it is allowlisted by destination per character rather than open by default.
  • Credentials stay outside the agent. The browsing service is credential-fronted — see Secrets & Egress for the general rule this follows.
  • It is a tool, not a crawler. Forge browses pages an agent has a reason to open. It is not a site-wide scraping engine and does not try to be.

Under the hood

Ready to ship faster?

Forge runs the full cycle autonomously — from labeled issue to reviewable PR.