Browser Task Design
Start With the Outcome
Tell the agent what result you need rather than prescribing clicks or selectors. A strong task names the target, required fields, relevant Profile, and actions that are forbidden or require approval.Request Verifiable Results
Ask for source URLs and explicit null values when information is missing. For time-sensitive research, request the date checked. This makes it possible to distinguish what the website stated from what the agent inferred.Keep Boundaries Explicit
For read-only work, say so. For actions such as submitting a form, posting, buying, applying, sending a message, or deleting data, state whether the agent may perform the action or must stop for approval. Authentication is a separate boundary. Name the Profile to use and ask the agent to pause for a direct browser sign-in rather than requesting secrets in chat.Reuse One Session
Related pages should normally be inspected inside one named Session. This preserves navigation state, keeps authentication scoped correctly, and makes the browser trail easier to inspect.Verify Before Finishing
The agent should verify the current URL, visible page state, and requested fields after navigation or an action. If remembered site structure conflicts with the live website, the live page wins and useful memory is refreshed.Handoff Checklist
A completed browser task should include:- The requested result in a consistent shape.
- A source URL for each item or claim.
- Clear nulls or caveats for unavailable information.
- A note about any authentication or approval boundary encountered.
- Confirmation that prohibited actions were not taken.