Guardrails

Guardrails enforce operational boundaries for agents. They help teams keep autonomy useful without sacrificing security, cost control, or compliance.

Guardrail Controls

• Tool approval gates (human-in-the-loop flows).
• Cost and token budget constraints.
• Action constraints by identity/role.
• Runtime policy checks before sensitive operations.

Approval-Aware Workflows

For high-impact actions, agents can request approval before tool execution. This creates a clear review point while preserving automation for low-risk steps.

Cost and Resource Protection

Budget-oriented guardrails reduce surprise spend by setting explicit limits and alerting behavior when thresholds are reached.

Best Practices

• Apply strict guardrails in production-facing environments.
• Use lighter guardrails in development with clear escalation paths.
• Separate approval policies by team and workload risk.
• Keep policy definitions in agentspecs so behavior is reproducible.

Next Steps

• Attach policy to config in Agentspecs.
• Review runtime behavior with Outputs.

See also

• Tool Approvals
• Identities
• Monitoring