AgentDesk
A control-plane exploration for policy-aware computer-use agents.
System flow
How the pieces move.
- Task Declare the intended outcome.
- Policy Check boundaries and approvals.
- Browser Execute the allowed interaction.
- Trace Capture evidence for review.
Problem Statement
Computer-use agents can operate real interfaces, which makes policy boundaries, approval gates, trajectory capture, and deterministic evaluation essential.
Architecture Summary
Declarative tasks move through planning, policy checks, browser execution, trajectory recording, cost limits, and replayable evaluation reports.
Key Design Tradeoffs
- Uses a deterministic planner before attempting unrestricted computer control.
- Prioritizes replayable evidence over autonomy theater.
- Treats policy denials as expected system behavior, not exceptional failures.
Technologies
TypeScriptPlaywrightJSON SchemaPolicy engine
AgentDesk explores the enclosure around a computer-use agent: policy, approvals, traces, evaluation, cost controls, and evidence that can be reviewed after the run.