The Auto-Research Loop
Skills that improve themselves against a number.
An autonomous optimisation loop where an AI agent iterates on a single target, a prompt, module or Skill, to maximise a scalar metric computed by an automated evaluator. Hypothesise, modify, evaluate, keep if better else reset, repeat. The binding constraint is not the loop; it is knowing what to measure.
How it runs
A three-file architecture separates the goal (read-only), the target being optimised (read-write) and the evaluator that returns a single number the agent never sees the source of. The loop iterates until the metric stops improving. It is the optimisation engine that keeps a Skill library from going stale, every Skill improves continuously against its metric.
The hard part is metric design. As the source puts it: the skill of the future is knowing what to measure.
The field notes
Frameworks, builds and what is changing in AI search.
Sent occasionally, never noise. The thinking behind the work, and the experiments before they ship.
No spam. Unsubscribe anytime.