Skip to content
fastbrowsefastbrowse

How the agent acts

When a browser agent should stop, not act

fastbrowse stops before actions classified as irreversible, ending needs_confirmation with exit 3. What that default protects, and when to use --authorize.

verified against v0.12.0 on this post’s sourceall posts
Where a run stopsAn action classified as reversible is taken. One classified as irreversible ends the run at needs_confirmation with exit 3 before it is dispatched, and acts only when the run was started with --authorize.ONE ACTIONNext actionsubmit, pay, delete, sendClassifysafety rulesactfill, scroll, navigatestoppedneeds_confirmation, exit 3actrun started with --authorize

swipe the diagram sideways for the rest

Each action is classified before it is dispatched. A reversible change goes ahead; an irreversible one ends the run at needs_confirmation with exit 3, unless the run itself was started with --authorize.

A browser agent needs to distinguish a price lookup from a purchase. fastbrowse stops before actions classified as irreversible, such as submitting an order, paying, deleting data or sending a message, unless the run is authorized. It can still fill fields and make reversible changes. Classification can miss a change, so state read-only constraints in your task and check important actions yourself.

The default is to stop

When the next action is classified as irreversible, an unauthorized run ends with the status needs_confirmation and exit 3 before dispatching that action. A script that tests for a non-zero exit sees the stop without parsing the answer.

StatusExitWhat it means
complete0Jev judged the task done
needs_confirmation3stopped before an irreversible action; re-run with --authorize
needs_login4a sign-in wall that no --secret covers
blocked6a bot check that did not clear
budget_exceeded7a resource limit was reached

Only complete exits 0. The full table, including the statuses that stop before anything is typed, is in run statuses and exit codes.

Opting in, one run at a time

--authorize allows submit, pay, delete and send. It does not override a task that says to change nothing or not to download, so a task carrying its own constraint is still refused.

export SAUCE_PASSWORD=secret_sauce
fastbrowse "Log in as standard_user with the saved password and add the backpack to the cart." \
  --start https://www.saucedemo.com/ --secret password=SAUCE_PASSWORD --authorize

Without --authorize, the run stops if an action is classified as irreversible. The exact stopping point depends on that classification. Which flags the run accepts is in the command-line reference.

The same default behind an MCP server

fastbrowse-mcp is started without --allow-authorize, so every call that reaches an irreversible step stops at needs_confirmation and the calling model cannot opt itself in. With the flag on, and --max-steps, --max-dollars and --max-seconds set, you know the most a call can do and the most it can cost before the first one runs. Setting those limits server-side covers the flags.

What to branch on in your own code

  1. Stop the loop on exit 3 rather than retrying it. A retry spends another run to arrive at the same refusal.
  2. Inspect the stopped result and let a person complete the step, or re-run with --authorize once you have decided the task may complete on its own.
  3. Keep --json. It carries the status, the step count, the cost and each quote behind the answer, which is what an audit of the run reads.
uvx fastbrowse "what is the current price of the pro plan" --start https://example.com/pricing

Next: your first run.

hosted

Run fastbrowse over an API

The same agent behind an API. Keys, secrets, run history and spend limits live in the console. Free to start.

create an account·plans and pricing