Skip to main content
Desktop tasks evaluate an agent’s ability to interact with full operating system environments—manipulating files, using applications like Excel, Word, and browsers, and executing multi-step workflows.

Task Components

1. Prompt

A natural language description of the goal. Example: “Extract and analyze quarterly sales data from the business report. Open ‘Q2_2025_Business_Report.docx’ in Word and review the sales data. Then, create an analysis spreadsheet in Excel…“

2. Initialization Config

Programmatic setup that ensures consistent starting state:

3. Evaluators

Verifiers that determine completion success via:
  • File state comparison
  • Application state checking
  • Screenshot analysis

Initialization Config Structure

Initialization Functions


Evaluator Configuration

Desktop evaluators compare the agent’s output against gold standard files:
See Desktop Verifiers for full evaluator documentation.

OS-Specific Considerations

macOS Example (Save before evaluation)


Best Practices

Ensure every agent run starts from the exact same state. Download fresh files rather than relying on pre-existing VM state.
Focus on verifying the end result (file content, structure) rather than the exact steps taken. Multiple valid paths may exist.
Most applications don’t auto-save. Always include explicit save commands before running evaluators.
Application behavior varies by OS. Test your gold standard and evaluator on the same OS the agent will use.
Choose specific evaluators (pivot_table, compare_docx_lines) over generic ones (exact_match) for meaningful comparison.

Next Steps

Desktop Verifiers

Configure file comparison and rules-based evaluation

MCP Tasks

Tool-use tasks with trajectory verification