agentx-harness · 1.0.7

Benchmark your own chips with AgentX

Install agentx-harness, point it at your inference server, and replay a public agentic workload on your own hardware.

Run AgentX

Start an OpenAI-compatible server, then set your model, URL, and concurrent session count below. Requires Python 3.11–3.13; 16 sessions is just an example.

Run AgentX

pip install 'agentx-harness==1.0.7'

export HUGGINGFACE_MODEL_NAME="org/model"
export CONC_SESSIONS=16

agentx profile \
  --url "http://127.0.0.1:8000" --endpoint /v1/chat/completions \
  --model "$HUGGINGFACE_MODEL_NAME" --tokenizer "$HUGGINGFACE_MODEL_NAME" \
  --tokenizer-trust-remote-code \
  --concurrency "$CONC_SESSIONS" \
  --public-dataset "semianalysis_cc_traces_weka_062126" \
  --output-artifact-dir "output_dir_results"

Using a model alias? Change --model and keep --tokenizer set to the matching Hugging Face tokenizer. Use --tokenizer-trust-remote-code only for trusted repositories. Omit it if custom tokenizer code is unnecessary.

Optional routing headers

Export these before running AgentX if your router needs session headers. Only the Dynamo option sends the parent-session ID for subagents.

Optional routing headers

# Adds X-Dynamo-Session-ID; subagents also get X-Dynamo-Parent-Session-ID.
# Only this option sends the parent ID, preserving forked-agent lineage.
export AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID=true
# Also sends X-Session-ID, if a front-end router needs it.
export AIPERF_HTTP_X_SESSION_ID_FROM_CORRELATION_ID=true