Custom AI Agent with Real-Time Corules Constraint Enforcement

Das Problem

Integrate Corules as middleware for homegrown AI agents. Gate 1 bounds the agent's action space before reasoning. Gate 2 validates the final decision before execution. Supports Python and Node.js SDKs.

Einrichtungsschritte

1. 1

   Install guardrail-sdk: pip install guardrail-sdk (Python) or npm install @guardrail/sdk (Node)

2. 2

   Initialize client with tenant_id and API key from environment variables

3. 3

   Before agent reasoning: call client.get_constraints(use_case, actor, context) — pass constraints to agent system prompt

4. 4

   After agent produces decision: call client.validate(use_case, actor, context, decision) before execution

5. 5

   Handle outcomes: ALLOW → execute, ESCALATE → route to human review queue, BLOCK → return error to agent

Codebeispiel

# Python: two-gate enforcement pattern
from guardrail import CorulesClient

client = CorulesClient(
    tenant_id=os.environ["GUARDRAIL_TENANT_ID"],
    api_key=os.environ["GUARDRAIL_API_KEY"],
)

# Gate 1: get constraints before agent reasons
constraints = client.get_constraints(
    use_case="discount_approval",
    actor=actor_jwt,  # signed JWT — never trust self-reported identity
    context={
        "customer_tier": opportunity["customer_tier"],
        "deal_value": opportunity["amount"],
        "product_category": opportunity["product_category"],
    }
)

# Pass constraints to agent
system_prompt = f"""
You are a sales assistant. Propose a discount for this opportunity.
Constraints:
- Maximum discount: {constraints['max_discount_pct'] * 100}%
- Required approval above: {constraints.get('escalation_threshold_pct', 0) * 100}%
"""

# ... agent reasons and produces proposed_discount ...

# Gate 2: validate before executing
result = client.validate(
    use_case="discount_approval",
    actor=actor_jwt,
    context=context,
    decision={"discount_pct": proposed_discount}
)

if result.outcome == "ALLOW":
    execute_discount(opportunity, proposed_discount)
elif result.outcome == "ESCALATE":
    route_to_human_review(result.audit_id, result.escalation_context)
else:  # BLOCK
    raise PolicyViolationError(result.violations)

Frequently Asked Questions