Axiom.ai v5.4: The Browser Agent asks less and does more, Code Tool session settings, and more

Thank you to everyone who has tried the new Browser Agent. Your feedback has helped us shape and improve the tool in this release. We have fixed a number of issues that stopped the agent completing a build, and added new capabilities. This is the first of many releases that will keep improving the agent.

The Browser Agent asks less and does more, Code Tool session settings, and more

Improved agent onboarding


A new design for the Browser Agent. The agent now asks fewer questions and lets you get working more quickly. It now asks for sheet and webhook information when it needs it rather than as a pre-qualifying stage, and the same start screen lets you move straight into a manual build if you prefer to place the steps yourself.

The New automation screen: the Build with Browser Agent card at the top, then the Add step box and the step list for building step by step

The agent is now able to use the existing session on your PC without storing it, so you can build your axiom without having to save login credentials. Stored cookies are still required for API or cloud scheduled runs.

Agent capabilities enhanced


  • The agent can now enable bot block bypass and iframe functionality when building an automation
  • Pager handling enhanced, with more options for specifying page numbers
  • Improved handling of the flow when data needs to be mapped into an automation that fills in a form
  • Agent automation naming improved
  • Error handling in loops improved to show more clearly what went wrong at a particular step
  • Improved automation handoff, eliminating cases of confusing or contradictory wording once the agent has finished its task and is handing back to the builder
  • Answering "yes" or "ok" in your own words now counts as agreeing, so the agent doesn't ask the same question multiple times
  • Form submissions are by default not triggered in verification runs
  • Improved loop handling; the agent no longer has to rebuild the entire loop when a single edit needs to be made

The agent at work: the live browser on a BBC search page, and the AI Assistant panel asking which of the headline, link, description and published time to write to the Google Sheet

More settings for Code Tool sessions


In the Code Tool, browserOpen() takes the same session settings as an automation: iframes, bypassBotDetection, timezone, blockResources and proxy. A mistyped option now throws instead of being ignored. The Claude MCP server's open_browser takes the same settings, and can open a session set up exactly like one of your automations.

Step API sessions opened with an API key and recording switched on now have their replay video, DOM log and console log on the code dashboard, and the Console Log tab shows the page's console output for Code Tool sessions.

Optional clicks run faster by default


The Click element step's Optional click setting is now a switch with a timeout beside it, 10 seconds by default. If the element hasn't appeared by then, the step is skipped without an error. Optional clicks used to wait 30 seconds, which made cookie banners and pop-ups that didn't appear slow down every run. Automations that already had the box ticked keep it on, with the new timeout.

Minor fixes


  • Fixed an issue where the Run button could stay on Starting… when the connection to the cloud browser never opened, or after a failed check at sign-in
  • Undo now returns to the correct state after an undo followed by an edit
  • Agent delete of the last step now no longer removes the add step button too
  • Fixed issues with token replacement in the builder clearing valid selectors
  • The step finder's shortcut no longer opens on top of the description form, menus and text fields
  • Delete my account in settings was slow; it has been sped up significantly
  • Signing in from the Your login session has expired dialog now loads correctly when on the builder and looking at a saved automation
  • The scheduler now accepts starting dates that are earlier on the same day. In these cases, the schedule will run as soon as it can.
  • Desktop scheduler no longer runs a minute late or skips runs when set to a one minute interval
  • Scheduled and API runs no longer lose a saved run that held only localstorage data
  • Fixed a hang in Click multiple elements failed when the set maximum was greater than the matches
  • Download files no longer fails with a cryptic error when running on the desktop app
  • Deleting axioms from the dashboard no longer leaves the deleted automation in the list for a few seconds

Previous