Lesson 07 · Activity 01

Agentic Browsing Brief

An agentic browser is only as safe as the instructions you hand it. A vague request — "find me some readings" — invites the agent to wander, click the wrong things, and act on whatever a page tells it. A brief does the opposite: it states the goal, draws the boundaries, and names the steps you must approve before they happen. This tool walks you through scoping one real task, then generates a portable brief you can paste into Comet, Atlas, or Auto-Browse.

Five decisions to settle first

Check each as you make it. Your progress is saved on this device.

0 of 5 done

Goal: can you state the task in one sentence, with a clear finish line? If not, it is probably two tasks.
Scope: which sites should it use — and which must it stay off? Naming the playing field keeps it from drifting.
Constraints: what rules shape an acceptable result (budget, date, source quality, format)?
Stop-points: which steps must it never take without your OK — buy, send, submit, log in, accept terms?
Data & privilege: is any sensitive data involved, and can the task run logged-out? Keep student records out entirely.

Brief builder

Fill in what you know. Empty fields are left out of the generated brief. Your entries are saved on this device until you change them.

The goal, in one sentence

What "done" looks like — the finish line

Sites it should use — optional

Sites it must avoid — optional

Constraints on an acceptable result

Stop-points — it must pause and ask before…

Buying / paying Sending a message Submitting a form Logging in Accepting terms Downloading files

Engine you'll hand it to

Privilege level

Fill in the fields above, then choose "Generate brief" to build your portable brief here.

Paste this brief as your opening message to the agent. Keep the stop-points line intact — it is what turns an autonomous run into a supervised one. Then watch it work and verify the result against the actual pages, not just the agent's summary.

Reflection

Which boundary was hardest to set — the scope, the constraints, or the stop-points — and did writing it change how comfortable you'd be delegating this task?

Saved automatically on this device.

Why the brief comes first: The difference between a useful agent and a dangerous one is rarely the model — it's the instructions. A one-sentence goal, a named playing field, and an explicit list of "ask me first" steps convert delegation from a leap of faith into a supervised task. If a task can't be bounded this cleanly, that's a signal to do it yourself.