Self-Improving Agent
the thing that acts
A system that rewrites everything about itself.
The loop
- 01 · Propose
- Choose the change: scaffold, recipe, or architecture.
- 02 · Spend
- Decide how much compute the change deserves.
- 03 · Keep
- Keep only what a held-out world says is better.
- 04 · Rewrite
- Point the same policy at the improver itself.
What it owns
The policy over changes: what to try, how much to spend, when to stop, what to keep. The scaffold it runs in, the recipe that trained it, and eventually the architecture underneath it are all just surfaces this policy edits.
Why it can improve itself
Every research loop today ends with a person deciding what to try next. That decision is the only step in the loop that compounds, and it is the last one anybody has automated.
An agent that improves its own scaffold is the one case where the thing being improved and the thing doing the improving are the same running system. That is what makes it recursion rather than tuning.
What it looks like when it works
A descendant improver produces more improvement than the ancestor it came from, on a target system neither has seen, at equal compute.