Controlled evolution
This direction explores improving an agent while a person controls changes. It is under development, not a promise of an available autonomous feature.
Section 06 / AIIn development
Agent development
Research into agents learning from results.
An Extella research direction: controlled evolution and agent sleep. The map marks these capabilities as work in progress.
Why it helps
The idea is to improve agent behaviour and knowledge from the results of its work, with changes kept under human control.
What it contains
Agent sleep
The map uses this name for work with accumulated experience and knowledge. It is also marked as work in progress.
What was measured
Extella measurements, September and October 2026. Each number holds under the conditions given.
- A night of training on a laptop.
Experiment of 30.09.2026: the small Qwen3-4B model on a Mac, an exam of 80 questions about Extella tools, worded differently from the training material. Without the knowledge the model answered 0 questions out of 80 correctly; with a cheat sheet in the request, 75 out of 80, and the cheat sheet took 400 tokens. After one night of training lasting 27 minutes, 70 out of 80 without the cheat sheet, with 30 tokens in the request.
- A person decides on switching on and rolling back.
A change switched on through the platform gave 15 correct answers out of 16; after rollback, 0 out of 16.
- This is research, not a finished feature.
One run on one Mac and one set of tasks. How it works on other models and tasks has not been tested.
Articles and demos
Related articles and examples.