Library
ha
Research · 2026
harness-arena
agent-harness evals through games
48 commits
A/B preset comparison design
arena run with preset A/B
win-rate by harness preset
About
A browser-first evaluation lab for tuning agent-harness presets through games: repeatable measurements of how harness settings change agent behavior.
Shelved while the trading systems take priority; the arena loop and preset comparison design are in place.
Liner notes
- Games as eval environments: legible and repeatable.
- Preset differences measured, not eyeballed.
Credits
Language typescript
Shape browser-first lab
Status shelved