Back
GitHub

BrowserSkill: Bounded Browser Agency

A practical model for scoped action, fresh evidence, and honest recovery

11 chapters103 min total
GitHub
Tech
Keywords
BrowserSkillagencysessionsAgent Windowtabsborrowingobservationauthorizationinterruptioncleanup
Overview

This series explains how BrowserSkill lets an AI agent use a real, already-logged-in browser through a layered bridge while keeping windows, tabs, permissions, human decisions, and uncertain page effects visible. It progresses from the CLI, daemon, and extension to session ownership, borrowing, observation, action, human handoff, interruption, operating boundaries, and cleanup. The result is bounded agency: useful browser control with explicit scope and evidence, not a guarantee of business success or transactionality.

Learning outcomes
  • BrowserSkill makes real-browser access governable through session scope, explicit tab claims, and visible cleanup.
  • A visible tab is not automatically authorized; remote targets require task creation or explicit borrowing.
  • Fresh observation connects page meaning to action, but browser results remain narrower than website business outcomes.
  • Human handoffs and interruptions preserve explicit boundaries and uncertainty instead of proving broader task success.
Keywords
BrowserSkillagencysessionsAgent Windowtabsborrowingobservationauthorizationinterruptioncleanup

Chapters

1
BrowserSkill Agent Window and Tab Ownership
How explicit borrowing keeps real-browser access inspectable
10 min · How explicit borrowing keeps real-browser access inspectable
10 min
2
BrowserSkill Request Routing Architecture
How the CLI, daemon, and extension preserve scope
10 min · How the CLI, daemon, and extension preserve scope
10 min
3
BrowserSkill Session Lifecycle
How session state governs startup, teardown, and recovery
8 min · How session state governs startup, teardown, and recovery
8 min
BrowserSkill Tab Claims and Borrowing
Why visibility is not authority across local and remote sessions
10 min · Why visibility is not authority across local and remote sessions
10 min
BrowserSkill From Intent to Authorized Action
A state model for authorization, effect, and cleanup
8 min · A state model for authorization, effect, and cleanup
8 min
BrowserSkill Semantic Observation
Fresh references for page meaning, evidence, and action
9 min · Fresh references for page meaning, evidence, and action
9 min
BrowserSkill Safe Actions from Fresh References
How geometry, native input, and verification limit claims
9 min · How geometry, native input, and verification limit claims
9 min
BrowserSkill Human Help and Borrow Approval
How confirmation and handoff outcomes limit authority
9 min · How confirmation and handoff outcomes limit authority
9 min
BrowserSkill Interruption and Unknown Effects
How queues, cancellation, reconnects, and sandboxes change diagnosis
9 min · How queues, cancellation, reconnects, and sandboxes change diagnosis
9 min
BrowserSkill Local, Remote, or DSH
How daemon location changes trust, evidence, and capability
11 min · How daemon location changes trust, evidence, and capability
11 min
BrowserSkill Adoption Boundaries
What the control loop can prove, and where it stops
10 min · What the control loop can prove, and where it stops
10 min

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans
Ch.1-3 FreeCh.4-11 Premium

FAQ

What is BrowserSkill?

BrowserSkill is a layered bridge that lets an AI agent request scoped work in a real browser while the user retains control of tabs, human decisions, and unresolved effects.

How does BrowserSkill prevent an agent from controlling the wrong tab?

It separates the session's Agent Window from ordinary windows and requires session ownership or an explicit borrow before page control; visibility alone is not authorization.

Why must an agent observe again after a page change or interruption?

It must observe again because navigation, rerendering, or an unknown effect can invalidate references and leave the current page state uncertain.

Does a successful BrowserSkill operation prove that a website task succeeded?

No, a successful BrowserSkill operation does not prove that a website task succeeded. It confirms only the narrower browser-level result reported by the tool; the website's business outcome may still be unconfirmed.