Skip to main content

What is Simthesizer?

Simthesizer (simulator + synthesizer) is an agent-driven simulation framework for LLM serving systems. It ships one simulator that evolves: a coding agent gets the representation and the process to extend it for the serving system you care about.

It has two parts:

  • The framework. Skills shipped in .agents/skills/ take a natural-language feature request through specification, mapping, implementation, and validation. Requirements stay traceable, and fidelity is checked against independent observations such as real runs. See Agent workflow.
  • The simulator. A deterministic discrete-event simulator in Rust expresses the complete serving workflow, control decisions included, as one dynamic event graph. New mechanisms enter as graph nodes owned by components, so one shared simulator evolves instead of being rewritten. See Simulator overview.

What it simulates​

The native input is an agent workload: each request is a DAG of LLM calls and tool calls, with think time between them. The output repeats that DAG and adds per-token latency to every LLM call.

Compute costs come from per-layer latency tables profiled with vLLM on real hardware. Communication costs come from a flow-level network model. Simulating needs no GPU.

Where to go next​