/ Engineering note · Synthetic load test

Load-testing an AI gateway

An engineering note on a controlled synthetic load test against an AI Gateway. More than 69,000 requests were generated by the test harness — synthetic traffic, not production traffic — to exercise the gateway path under sustained load. Framework and platform evaluation was run separately, as a qualitative architecture comparison.

/ Context

An AI Gateway sits on the request path of every application that uses it, so its behaviour under sustained load is a platform concern rather than an application one. The question was how the gateway path itself behaves when driven hard — separate from any question about which model or framework an application should use.

/ Role & decision scope

  • Designed and ran the synthetic load test against the AI Gateway.
  • Built the test harness that generated the request load.
  • Kept qualitative framework evaluation separate from the gateway load test.

/ Constraints

  • Synthetic traffic only — this was a controlled test, not observation of production usage.
  • The test exercises the gateway request path, not end-user application behaviour.
  • Frameworks and platforms differ in purpose, so they are not directly comparable as performance targets.

/ Architecture & key decisions

  1. D01

    Test the gateway, not the vendors

    The load test targets the AI Gateway request path. It is deliberately not a head-to-head performance ranking of OpenAI, Amazon Bedrock, LiteLLM, AgentKit, LangChain and AWS Strands — those serve different roles and are not equivalent targets.

  2. D02

    Framework evaluation kept qualitative and separate

    Those frameworks and platforms were assessed on architecture fit, integration surface and operability as a separate exercise, feeding the platform ADRs rather than a latency table.

  3. D03

    Synthetic load, stated as synthetic

    More than 69,000 requests were generated by the harness during the test. That is a property of the test run, not evidence of adoption, enterprise scale or production traffic.

/ Trade-offs

Synthetic load versus real traffic realism
A synthetic harness gives a repeatable, controllable signal about the gateway path; it does not substitute for production observability, and is not presented as such.
A single comparable number versus honest comparison
Non-equivalent platforms were compared qualitatively on architecture fit instead of being forced onto one benchmark axis.

/ Evidence & outcomes

Artifact
Load-test harness generating 69,000+ synthetic requests against the AI Gateway
Decision
Qualitative architecture evaluation of OpenAI, Amazon Bedrock, LiteLLM, AgentKit, LangChain and AWS Strands, recorded separately from load testing

/ Related capabilities

/ Next step

Let's talk about the work.

Open to full-time Senior–Staff AI / agent platform roles (remote-friendly), as well as contract and consulting engagements. Response within two business days.