Browser Agents Overview
Browser agents are autonomous bots that control a web browser to complete tasks described in natural language. Given a goal, an agent observes the current page (screenshot + HTML), reasons about what to do next, and takes action — click, type, navigate, extract. No scripts required.
How They Work
Each step, the agent:
- Captures a screenshot and reads page structure
- Consults the LLM with the current state and goal
- Decides on an action (click, type, navigate, extract)
- Executes the action and observes the result
- Repeats until the goal is complete or a failure is detected
The system prompt you write defines the agent's behavior and constraints.
Use Cases
- Automated form submission and account workflows
- Price and content monitoring
- Data extraction across multiple pages
- Regression testing after UI changes
Browser Agents vs E2E Tests
| E2E Tests | Browser Agents | |
|---|---|---|
| Execution | Follows a script | Decides in real-time |
| Selectors | CSS / XPath | Understands element purpose |
| UI changes | Breaks | Adapts |
| Best for | Stable critical paths | Complex flows, dynamic UIs |
Use E2E tests for CI gates and compliance paths. Use browser agents for exploratory testing, complex conditional flows, and workflows that change frequently.
Access
Browser Agents is currently in private beta. Access is granted on request — contact the team to enable it for your account.
Agent Lifecycle
| Phase | Status |
|---|---|
| Create, configure, run, view training | Available |
| Labeling UI, evaluations, deployment versioning, production monitoring | In development |
Agents start in draft status. Full deployment pipeline (versioning, channels, production monitoring) is coming.
Glossary
| Term | Meaning |
|---|---|
| System Prompt | Natural language instructions defining agent behavior |
| Run | Single execution of the agent |
| Step | One action within a run (click, type, navigate, extract) |
| Outcome | success / failure / error |
| Golden Dataset | Curated training examples |
| Labeling | Reviewing run steps and marking correct/incorrect actions |
| Training Run | A job that improves the agent from labeled examples |
| Version | Snapshot of the agent at a point in time |
| Channel | draft / candidate / deployed |
Next: Creating an Agent