Browser agents often begin with a broad instruction but lack the operator’s hidden knowledge: which path works, what success looks like, where exceptions occur, and which actions should never happen without approval.
Automate
AgentTwin
Teach the browser once. Let an agent handle the repeat.AgentTwin is TabTwin’s adaptive execution layer—not its only AI. It understands a goal, plans across specialized Twin capabilities, validates outcomes, and requests human approval or help at defined risk boundaries.
01
Why this capability exists
From friction to a clear outcome
AgentTwin starts from a demonstrated human workflow. The agent observes page state, reasons within defined rules, acts through approved browser primitives, validates outcomes, and escalates uncertainty or sensitive actions to a person.
02
Real-world scenario
Turn invoice emails into supervised ERP expense entries
- 1
Demonstrate opening an invoice email and extracting its fields
- 2
Show how the matching ERP expense record is created
- 3
Define duplicate checks, validation rules, and limits
- 4
Require human approval before the final posting step
- 5
Run the agent with logs, exceptions, and review controls
03
End-to-end workflow
How AgentTwin works
- 01
Describe the goal or demonstrate a successful workflow.
- 02
Let the shared AI Core route research, data, form, flow, and task work to their specialized Twins.
- 03
Run the adaptive plan with logs, exceptions, budgets, and approve/edit/reject controls for sensitive actions.
04
What is included
Core capabilities
Demonstration learning
Uses a successful human run as grounded operating context for the agent.
Observe–reason–act loop
Reads current page state, chooses an allowed step, acts, and checks the result.
Triggers & schedules
Starts from an event, queue, manual request, or approved schedule.
Validation rules
Defines success, duplicate checks, limits, required values, and stop conditions.
Human approvals
Pauses for approve, edit, reject, or takeover at sensitive boundaries.
Logs & exceptions
Records decisions, actions, outcomes, retries, and cases that need help.
05
Where it wins
Best use cases
- ✓Lead research and CRM creation
- ✓Routine operations across internal tools
- ✓Long-running browser processes with exceptions
06
What the user gets
Expected outputs
- →Human-supervised browser agent
- →Approval queue and exception handling
- →Auditable execution history
What changes for the team
- ↗Automate longer browser processes without hiding decisions
- ↗Transfer operator knowledge into a supervised repeatable system
- ↗Keep people focused on exceptions and high-risk choices
Safe by design
- ✓Domain, action, schedule, and spending limits constrain every run
- ✓Sensitive or irreversible steps require explicit human approval
- ✓Auditable logs and takeover controls remain available throughout execution
AgentTwin combines a real human demonstration with explicit permissions and validation, creating an agent that is grounded in how the work actually succeeds.
One shared technical foundation
Powered by the Browser Session Engine.
AgentTwin consumes the same structured DOM events, timeline, context, snapshots, and safe execution primitives as the rest of TabTwin—without requiring another extension.