Metlo’s agent is designed to be performant and reliable. We did this by writing all of our common agent logic in a shared rust library. This guarentees that all of our code is performant and memory safe. Each of our agent has bindings to this native library.

In addition to being written in rust, our agent asynchrously pulls models that specify how to block. This minimizes the additional latency added to each request.

To test each agent we send requests ranging from 1KB to 1MB at ~500rps for 24 hours against a test server. These are the maximum system metrics measured:

  • Max CPU Usage - 1%
  • Max Memory Usage - 50MB
  • Max Latency Increase - 0.2ms

In practice Metlo should use much less resources that this.