AgentDyno — the dyno readout

Success vs inference cost across harness architectures.

Static snapshot generated 2026-09-20 17:53 UTC from a real agentdyno run. Not a live app (Hugging Face's free Static Space tier can't run a Streamlit server) — for the live version run streamlit run dashboard/app.py locally, or see the canonical repo: github.com/muraliikrishnant/AgentDyno.

Pareto frontier: success rate vs cost per solved task

TTFT / ITL by trial

Single-trajectory drill-in