Policy Management

The CugaAgent includes a powerful policy system accessed via agent.policies. This allows you to enforce security guardrails, guide agent behavior with playbooks, and require human approval for sensitive actions.

Policies are triggered based on user input, context, or tool usage. You can use keyword triggers (exact matching) or natural language triggers (semantic matching) to determine when policies should activate.

Policy Types

Intent Guard

Blocks the agent from executing actions that match specific forbidden intents. Use keyword triggers or natural language triggers for flexible matching.

See Intent Guard documentation for details.

Playbook

Provides guided workflows or standard operating procedures (SOPs) to the agent when specific topics are triggered. Supports automatic refinement based on conversation history.

See Playbook documentation for details.

Tool Enrichment

Automatically injects additional context or instructions into tool descriptions when specific criteria are met.

See Tool Enrichment documentation for details.

Tool Approval

Requires human confirmation before executing specified sensitive tools. Implements human-in-the-loop workflows.

See Tool Approval documentation for details.

Output Formatter

Formats agent responses based on triggers and configuration. Transforms or restructures the final output before it's returned to the user.

See Output Formatter documentation for details.

Trigger Types

Policies can be triggered using different methods. See the Triggers documentation for detailed information about:

• Keyword Triggers: Exact word/phrase matching
• Natural Language Triggers: Semantic similarity matching using LLM

Management API

list

List all configured policies.

policies = await agent.policies.list()
# Returns List[Dict]

get / delete

Manage policies by ID.

# Get
policy = await agent.policies.get("policy_id_123")

# Delete
success = await agent.policies.delete("policy_id_123")

load_from_json

Load policies in bulk from a JSON file.

result = await agent.policies.load_from_json("policies.json")