Experiments
The Lab
Questions worth exploring, even when I don't know the answer yet.
This is where I document experiments, prototypes, and rabbit holes. Some will turn into projects, some will end, and some may just make me a little smarter. That's the point.
Evolving an Agent From Three Tools
Can a fixed model build a better harness around itself?
The design behind G000: three primitive tools, a mutable scaffold, and independent evaluation before any claim of improvement.
What G000 Calibration Actually Shows
What can I conclude from the pilot, low control, and GLM reference?
Low scored 18/24, the mixed DeepSeek pilot scored 20/24, and the GLM reference scored 21/24. Their run profiles limit what those scores establish.
What Building Ledger Eval Taught Me
Does the benchmark catch the mistakes I think it catches?
Exact-cent invariants, ambiguous reconciliation rules, incomplete refactors, and transport failures made the evaluator part of the experiment.
Featured
Evolving an Agent From Three Tools
ActiveCan a fixed model build a better harness around itself?
I've spent a lot of time building machinery around coding agents. This experiment starts from the other end: `read`, `write`, `bash`, and a loop. I want to find out whether a fixed model can propose useful changes to that scaffold, and whether an independent evaluator can show that those changes helped.
Read the full entry→$ lab status evolving-agent
- status
- active
- started
- Sep 2026
- updated
- Oct 2026
- tags
- agents, evals, self-improvement
▍
“The point of the Lab isn't to be right. It's to stay curious.”