Browser agents: automating work in apps that have no API
A browser agent uses a website the way a person does: it clicks, types and reads the page. When it helps, what can go wrong and how to keep it safe and checkable.
By the Zoft team · · 2 min read
Much real work happens in web applications that have no API, or whose API does not cover what you need: a supplier portal, an insurance site, a government form, an old internal tool. A browser agent is an AI agent that works in those pages the way a person would. It opens the site, reads what is on screen, clicks, types and moves on.
How a browser agent works
- It is given a bounded task, for example: download this month's invoice from the supplier portal.
- It opens a browser session and looks at the page.
- It decides the next action, such as clicking a link or filling a field, and does it.
- It checks the result, and repeats until the task is done or it gets stuck.
- It returns the outcome together with evidence of what it did.
When a browser agent is the right choice
- The system has no API, or the API is too limited.
- The task is repetitive: the same steps across many records.
- The task is bounded and has a clear finish: it is done when a file is downloaded or a form is submitted.
- Getting a human to do it is mainly a matter of time, not judgement.
When to prefer an API or a person
If a reliable API exists, use it. It is faster and less brittle than driving a page. And if the task involves a risky decision, such as payments, deleting data or legal submissions, put a person at the key step.
What can go wrong, and how to limit it
- Pages change. Buttons move and a login flow gains a new step. Build in checks that fail loudly instead of guessing.
- Overreach. Give the agent a bounded task and limited access, not a whole account.
- Credentials. Store them in a managed connection, never inside instructions.
- Unverified results. Require evidence, such as a screenshot or a downloaded file, so a person can confirm the work.
Browser agents in Zoft
Zoft's browser agents complete bounded work in live web applications with evidence. They run on the same canvas as workflows and voice and chat agents, so a call can trigger a browser task and the result can flow back into your systems. Shared evals and controls cover permissions, approvals and run records. To compare agents with fixed automation, read AI agents vs workflow automation.
Frequently asked questions
- What is a browser agent?
- A browser agent is an AI agent that works in web pages the way a person does: it opens a site, reads the page, clicks and types, and returns the result with evidence of what it did.
- When should I use a browser agent instead of an API?
- Use a browser agent when the system has no API or the API does not cover the task, and the work is repetitive and bounded. If a reliable API exists, use that instead.
- How do I keep a browser agent safe?
- Give it a bounded task and limited access, store credentials in a managed connection, require evidence for important steps, keep a record of each run and place a person at risky steps.