STAC agentic quant research benchmark launch

Powered by SwarmSimulator.

STAC's first Agentic POC did not invent prompts. SwarmSim captured a live 300-turn quant-research investigation from a US market maker, then replayed independent copies as the load. The industry can now score a serving stack on work a real desk already ran.

This is not chatbot Q&A. An agentic coding system was directed to train and evaluate trading models against market data. The elapsed time of that investigation now sits with the inference stack.

300

Sequential turns

172k

Median input tokens

1 to 64

Concurrent copies

Three hundred sequential turns. Median input 172k tokens. Context that accumulates, then compact. Turns that cannot be overlapped. How fast the stack generates each response, and how it carries context into the next request: that is what STAC timed.

STAC did not write synthetic prompts to scale the test. Load is scaled by replaying independent copies of the same captured work. SwarmSim perturbs each copy so the content stays semantically identical while the token sequence is new. No two copies collapse into the same cached state. That is how you go from 1 to 64 concurrent investigations without the benchmark cheating.

SwarmSim was the request engine, used as supplied: turn by turn, streamed completions, same prompt sizes on every stack. STAC's job was to measure the infrastructure. SwarmSim's job was to make the traffic honest.

If you run agents that think for hundreds of turns, chatbot scores will lie to you. The industry now has a way to score the stack on work a real desk already ran. That is what SwarmSim is for.

SwarmOptimizer is a different product. It was not the stack under test. STAC measured a Lambda 16xB200 cluster. SwarmSim is how the traffic got there.

Your agents can be next.

SwarmSim captures live work and replays it as load the cache cannot cheat. That is how STAC measured this POC. Talk to us about running yours.

Talk to Us Read the STAC report