Load Testing AI Safety Infrastructure

Safety infrastructure must keep pace with the agent traffic it protects. If the policy engine becomes a bottleneck under load, agents may experience timeouts or, worse, safety checks may be skipped. Load testing identifies performance limits before they cause production incidents.

What to Load Test

Policy evaluation throughput: How many envelopes per second can the policy engine evaluate? At what point does latency degrade?

Content scanning throughput: How many inputs per second can Aegis scan? How does scan time scale with input size under load?

Control plane API: How many concurrent requests can the control plane handle? What is the P99 latency under peak load?

Audit trail writes: How many receipts per second can be written to the database? Does write latency affect the critical path?

Test Design

Generate realistic load patterns. Use action envelope distributions that match production traffic: the same mix of action types, resource patterns, and parameter sizes. Synthetic load that does not resemble real traffic will not surface the same bottlenecks.

// Load test configuration
const config = {
  duration: '10m',
  stages: [
    { target: 100, duration: '2m' },   // Ramp up
    { target: 500, duration: '5m' },   // Sustained peak
    { target: 0, duration: '3m' },     // Ramp down
  ],
  envelope_distribution: {
    'search.web': 0.40,
    'data.read': 0.30,
    'data.write': 0.15,
    'file.upload': 0.10,
    'payment.send': 0.05,
  },
};

Key Metrics

Track during load tests:

• Throughput: Requests per second successfully processed
• Latency: P50, P95, P99 response times
• Error rate: Proportion of requests that fail or time out
• Resource utilization: CPU, memory, database connections, and network I/O

Finding the Breaking Point

Gradually increase load until the system degrades. The breaking point is where latency or error rate exceeds acceptable thresholds. Document this as the system's capacity limit. Plan production capacity to stay below 70% of this limit under normal conditions.

Safety Under Load

The most important question in load testing safety infrastructure: does the system fail closed under extreme load? When the policy engine is overwhelmed, are actions denied or are they allowed without evaluation? Verify this explicitly.

Regular Testing

Run load tests after every significant change to safety infrastructure and before scaling agent deployments. Traffic patterns change over time, and yesterday's capacity may not be sufficient for tomorrow's workload.

Load testing is not optional for safety infrastructure. It is how you know your safety holds under pressure.