Platform · Module
Browser & computer-use agent
Operate portals with no API in a sandbox, with screenshots and an approval gate
The browser agent operates legacy and government portals that have no API inside an isolated sandboxed browser. Credentials stay in the secret store, every step is screenshotted, and submissions always need approval.
For portals without an API
Some important work still happens on portals with no API: filing preparation, downloads and status checks. Planned for Wave 4 and metered per run, the browser agent operates these portals so people do not have to.
Isolated and observable
Each task runs in its own sandboxed browser, isolated with no persistent cookies across tenants. A task definition sets the goal, allowed domains, allowed steps, credential references, stop conditions and budget. Every step is logged with screenshots and DOM action records.
Credentials never reach the model
Credentials come from the secret store by reference and are never exposed to the model.
Approval before anything final
An approval gate sits before every submit, pay or sign step, with a preview of the final screen. Submissions are never autonomous, whatever the agent's autonomy level. Read-only tasks, such as status checks and downloads, run with health checks.
Reliability
A recorded human demonstration can become a task template after review. Portal change detection pauses a task when the layout drifts. CAPTCHA and OTP steps are handed to a human and never bypassed. Each portal domain's terms of use are gated through the Fact Register.
How it fits
The browser agent is a last resort, used only where no API or licensed connector exists. Integration Hub remains the preferred path. Every browser run is an Action with its own budget, trace and audit, so an admin can see exactly what was done on which portal and when. Downloads land as files in the tenant's storage and can feed the owning app, for example a statement into Finance.
Features
What Browser & computer-use agent includes
Sandboxed browser per task
Isolated runtime with no persistent cookies across tenants. W4.
Task definition
Goal, allowed domains, steps, credential refs, stop conditions and budget. W4.
Step log with screenshots
Screenshot audit and DOM action records. W4.
Approval gate
Before submit, pay or sign, with final-screen preview. W4.
Credential references
From the secret store; never exposed to the model. W4.
Read-only tasks
Status checks and downloads with health checks. W4.
Demonstration to template
Recorded human run becomes a reviewed template. W4.
Portal drift detection
Layout changes pause the task. W4.
CAPTCHA and OTP handoff
Handed to a human; never bypassed. W4.
Portal terms gating
Terms of use per domain checked in the Fact Register. W4.
Workflow
Download a monthly statement from a portal
- 1
Admin defines the task with allowed domain and credential ref.
- 2
The agent signs in inside a sandbox.
- 3
An OTP prompt is handed to a human.
- 4
The statement is downloaded and stored with screenshots of each step.
AI
AI skills and Digital Employees
Sandboxed portal operation
Navigate, read and download; submit, pay and sign always approval-gated.
Demonstration learning and drift detection
Templates from human demos; tasks pause on layout change.
Digital Employees that use this module
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Can the browser agent submit forms on its own?
No. Submit, pay and sign steps always need human approval with a final-screen preview.
Does it bypass CAPTCHAs?
No. CAPTCHA and OTP steps are handed to a human.
When is it planned?
Wave 4 (Year 3), metered per run.
See Browser & computer-use agent in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.