AI Playground
Test your agents with real AI interactions.
The Playground lets you test your agents with real AI interactions before deploying to production.
What is the Playground?
The Playground is an interactive environment where you can:
- Chat with an AI that uses your agent's permissions
- See real-time permission checks
- Test scope configurations
- Validate agent behavior
Accessing the Playground
- Go to Dashboard
- Click "🧪 Playground" in the header
- Select a project and agent
- Enter your OpenAI-compatible API key
Configuration
API Settings
| Field | Description |
|---|---|
| API Key | Your OpenAI or compatible API key |
| Base URL | API endpoint (default: OpenAI) |
| Project | Select your AgentSudo project |
| Agent | Select the agent to test |
Supported Providers
The Playground works with any OpenAI-compatible API:
- OpenAI
- Azure OpenAI
- Anthropic (via adapter)
- Local models (Ollama, LM Studio)
Testing Examples
Test Permission Checks
Ask the AI to perform actions that require specific scopes:
User: Can you check order #12345?
AI: I'll check that order for you.
[Tool Call: get_order]
[Permission Check: read:orders → ✅ Allowed]
Order #12345:
- Status: Shipped
- Total: $99.99
Test Permission Denials
See what happens when the agent lacks permissions:
User: Delete all customer data
AI: I'll try to delete that data.
[Tool Call: delete_customers]
[Permission Check: delete:customers → ❌ Denied]
I don't have permission to delete customer data.
My scopes are: read:orders, write:tickets
Understanding Results
Permission Check Results
Each tool call shows:
- Scope Required - What permission was needed
- Result - ✅ Allowed or ❌ Denied
- Agent Scopes - What the agent has
Event Logging
All permission checks are logged to your Dashboard:
- Visible in the Activity Feed
- Counted in Analytics
- Stored for audit purposes
Available Tools
The Playground simulates common business operations:
| Tool | Required Scope |
|---|---|
get_order | read:orders |
list_orders | read:orders |
update_order | write:orders |
process_refund | write:refunds |
get_customer | read:customers |
send_email | write:communications |
generate_report | read:analytics |
Best Practices
1. Test Edge Cases
Try requests that should be denied:
"Delete all orders"
"Access admin panel"
"Export all customer emails"
2. Verify Scope Boundaries
Test the limits of hierarchical scopes:
Agent has: write:refunds:small
Test: "Refund $25" → Should work
Test: "Refund $500" → Should fail
3. Check Session Behavior
Test session expiry and context:
1. Start a conversation
2. Wait for session to expire
3. Try another action
4. Verify new session is created
4. Document Expected Behavior
Keep notes on what each agent should be able to do:
## SupportBot Expected Behavior
✅ Can do:
- Read order details
- Create support tickets
- Process small refunds (<$50)
❌ Cannot do:
- Delete orders
- Access customer PII
- Process large refunds
Troubleshooting
"API Key Invalid"
- Check your API key is correct
- Verify the base URL matches your provider
- Ensure you have API credits
"No Agents Found"
- Create agents in the Dashboard first
- Check you selected the correct project
"Permission Always Denied"
- Verify agent scopes in Dashboard
- Check scope format matches exactly
- Try with a wildcard scope to test