003 / Kestrel
An agent that learns
Hidden rules learned from experience alone — no corpus, no pretraining — from tabular world models to self-modifying exploration, across six phases.
Write-up in progress.
01What it is
Kestrel is an attempt to build an agent that has to work out the rules of its world the way anything else does: by acting, being wrong, and revising. No corpus to learn from, no pretrained starting point, no hand-written reward for the thing that matters.
It moved through six phases, beginning with a tabular world model and ending with exploration that can modify its own search.
02Why 'no corpus' is the hard part
Most of what gets called agent learning is supervised learning wearing a loop. Give a model a corpus of solved problems and it will generalise from it, which is useful but is not the same problem. Take the corpus away and the sample budget becomes the whole difficulty: every rule costs experience, and experience costs time.
A tabular world model is a defensible place to start for exactly that reason. It is cheap, it is legible, and when it is wrong you can see which cell is wrong rather than guessing at a high-dimensional excuse.