I build open tools for people who run AI agents all day, mostly the part nobody enjoys: proving the work actually happened.
An agent that reports success is not evidence of success. Most of what I publish sits somewhere around that problem, from writing a finish line down before the work starts, to deciding whether the result really meets it.
What I'm working on
- hermes-jev-north-star - turn an intention into a checkable finish line, generate the run prompt from it, then let a judge decide whether the work is done. Deterministic checks first, an evidence file, then Jev for the parts a script cannot read. 82 tests, no key needed for the deterministic half.
- awesome-agent-verification - the list of tools and practices for deciding whether autonomous agent work is actually done. Eval harnesses, judges, tracing, guardrails, benchmarks.
- hermes-pets-all - a Hermes desktop plugin that browses every pet installed for a profile, past the built-in render cap.
How I work
Small tools that do one thing and say plainly when they cannot. If a check cannot be run, I would rather it fails than guesses. No telemetry, no accounts, no keys where a plain script will do.
Most of what I build is for Hermes Agent, the open-source agent from Nous Research. Everything public here is MIT or CC0 unless the repo says otherwise.