Skip to main content

What is Agent Execution?

Agent execution is how Granite runs your automations. An AI agent interprets your task, plans the steps, and executes them on a Windows machine.

Key Features

Execution Flow

1

Job Queued

When you run a process, a job enters the queue
2

Driver Assigned

An available driver machine claims the job
3

Connection Established

WebSocket connects dashboard to driver for real-time updates
4

Agent Starts

AI agent begins interpreting your task
5

Actions Execute

Each action:
  • Shown to you (with HITL)
  • Screenshot captured
  • Executed on desktop
  • Result verified
6

Completion

Task finishes, recording saved, results available

Agent Capabilities

The agent can:
ActionExamples
ClickButtons, menus, links, icons
TypeText fields, documents, search boxes
NavigateOpen apps, switch windows, scroll
ReadExtract text, verify content
WaitFor elements, for loading
DecideHandle variations, recover from errors

Execution Modes

Human-in-the-Loop enabled (default)
  • Agent pauses before sensitive actions
  • You approve, modify, or cancel
  • Maximum control and oversight

Real-Time Monitoring

During execution, you see:
  • Live video - The Windows desktop in real-time
  • Chat log - Each action described
  • Progress - Current stage indicator
  • Controls - Pause, cancel buttons

Error Handling

When something goes wrong:
  1. Agent describes the error
  2. Screenshot shows the state
  3. You can:
    • Modify the approach
    • Skip the action
    • Cancel execution

After Execution

Once complete, you have:
  • Execution summary - Success/failure, duration
  • Screenshot gallery - All captured images
  • Video recording - Full session replay
  • Detailed logs - Timestamped trace

Best Practices

Until you’re confident the automation works, keep HITL on.
Even successful runs may reveal optimization opportunities.
The clearer your description, the better the agent performs.
Run with HITL several times before disabling oversight.

Next Steps