The AI Agent Task Boundary Designer

 Build an interactive design studio for defining exactly what an AI agent is allowed to do during a multi-step task.

Users describe an agent's purpose, available tools, accessible data, permitted actions, and business constraints. The application converts these into a structured capability specification.

Implement:

1. Capability Matrix: Separate read, analyze, draft, modify, publish, delete, purchase, and administrative actions.

2. Action Boundaries: Define allowed targets, resource scopes, execution limits, and prohibited operations.

3. Approval Gates: Require explicit human approval for configurable high-impact actions.

4. Permission Inheritance: Visualize how permissions flow from the agent to its tools and delegated subtasks.

5. Attack Simulation: Test attempts to exceed scope, manipulate instructions, reuse expired authorization, or access unrelated resources.

6. Policy Compiler: Export the specification as JSON and generate adapter templates for enforcing it in an application.

7. Audit Simulation: Replay proposed actions and show whether each should be allowed, denied, or paused for approval.

Include example agents for customer support, code review, and business reporting.

Enforce authorization in the execution layer, not merely in the user interface or prompt text. Default to least privilege, deny ambiguous actions, and treat external content as untrusted input. Clearly distinguish simulated enforcement from actual security controls.