Pre-built policy templates for common AI agent scenarios. Copy them, customize them, and align them to your own runtime before production rollout.
Start quickly with ready-to-use configurations, then tune thresholds to match the actual risk of each workflow.
Essential protections for any AI agent
max_iterations: 20 max_cost_usd: 5.00 require_approval_for: [delete, write]
Comprehensive guards for production deployments
max_iterations: 10 max_cost_usd: 1.00 no_pii_in_output: true require_approval_for: [delete, drop, truncate] audit_all_requests: true rate_limit: 100/hour
Relaxed policies for development and testing
max_iterations: 50 max_cost_usd: 10.00 warn_only: true log_all_decisions: true
Prevent runaway costs by setting spending limits per request or session.
max_cost_usd: 1.00 warn_at_percent: 80 block_at_percent: 100
Limit token usage to prevent expensive model calls.
max_tokens: 4000 max_output_tokens: 1000 model_whitelist: [gpt-3.5-turbo, claude-instant]
Detect and block requests that would expose personally identifiable information.
no_pii_in_output: true detect_patterns: [email, phone, ssn, credit_card] redact_or_block: block
Prevent bulk data extraction and suspicious export patterns.
max_records_per_request: 100 block_bulk_export: true audit_all_exports: true
Require approval for any destructive or irreversible operations.
require_approval_for: [delete, drop, truncate] approval_timeout_seconds: 300 min_approvers: 1
Enforce content guidelines for AI-generated outputs.
content_policy: standard block_categories: [hate, violence, self_harm, sexual] require_human_review_for: sensitive_topics
Use the current loop-guard schema to defer runaway iterations, repeated tool calls, timeout overruns, and cumulative cost spikes.
max_iterations: 10 timeout: 60s same_tool: 3x no_repeat_calls: true max_cost_usd: 0.50
Control request frequency to protect downstream services.
requests_per_minute: 60 burst_limit: 10 cooldown_seconds: 60
Choose a template
Pick the template that matches the risk profile of your use case.
Customize values
Adjust thresholds, approvals, and limits to fit your own runtime and operating rules.
Deploy carefully
Add constraints to your `/check` requests. For loop guards, either send runtime counters in `context` or keep the same `run_id` across checks so the backend can persist state.