Newsroom
Same screen, two approaches: watching a script and an agent hit the same change
What happens when a portal layout changes under a scripted bot, and under a computer-use agent, step by step.
November 10, 2026 · 2 min read
Here's a test we run in the first week of most deployments. We take a screen the customer's automation already uses, change something small, and watch what happens.
The example below comes from a synthetic claims portal we built for demos. No customer data was used, but the change is one we see all the time.
The task
Open a claim. Read the policy number and the claimed amount. Check the amount against the policy limit in a second system. If it's under the limit, approve it. If it's over, route it to a reviewer with a note.
The change
The portal team ships a redesign overnight. The "Claimed amount" field moves from the right column to a new "Financials" tab. The Approve button is renamed "Submit decision".
What the script does
The script was built to look for a field at a fixed position, with a fixed label, and to click a button with a fixed name.
- It opens the claim. That works.
- It looks for "Claimed amount" where it used to be, finds nothing, and either fails or reads the wrong field.
- If it fails, the run stops and the item goes to the exception queue. Someone raises a ticket, and a developer updates the selectors and re-tests.
- Until then, every claim that night goes to a person.
Nothing went wrong with the script. It did exactly what it was written to do.
What the agent does
The agent is given the same task, written as an instruction rather than as a sequence of clicks.
- It opens the claim and reads the screen. The amount isn't where it was last time, so it looks for it, the same way a person would.
- It sees a "Financials" tab, opens it, and finds the amount.
- It checks the limit in the second system, and finds a button labelled "Submit decision" where Approve used to be.
- Claims over the limit go to a reviewer with the amount, the limit and the reason. That threshold is set by the customer, not by the agent.
The run finishes. The log shows each step, with a screenshot of what the agent saw.
What this doesn't mean
It doesn't mean every bot should be replaced. If a screen hasn't changed in three years, the script on it is cheap and reliable, so leave it.
It also doesn't mean the agent is never wrong. That's why it starts read-only, in a sandbox, and why anything it isn't sure about goes to a person. We measure completion rate and review every handed-back item before a single write action goes live.
