Rigr

Agent evaluation for teams that can't afford to be wrong. Define what your agent must output, freeze the known-good results, catch regressions before your customers do.

CLIinit · test · freeze · compare
YAMLversion-controlled cases
PyPIpip install rigr
MITopen-source core
pip install rigr
rigr init && rigr test

The problem

You have agents in production. Every model update, prompt change, or retrieval tweak can break them silently. The existing eval tools test whether the model sounds good. That is the wrong axis. I do not care if my agent sounds helpful. I care whether it still calculates the refund correctly after the model swap. Different question, and the one that pages you at 2am.

The workflow

Rigr records expected behavior in version control, runs repeatable cases, and compares new results with a reviewed baseline. The important idea is treating agent behavior as something a team can inspect in a code review, not something evaluated only by watching a demo.

Rigr is the sibling of Lethe. Rigr catches regressions between versions. Lethe measures degradation inside a single long run. Same instinct, two time scales: agents fail where nobody is measuring.

Open source on github. Backstory: measuring whether your AI is getting better or worse →