Lab experiment
Self-Improving AI Agents
Exploring how agents improve across attempts through execution feedback, verification, memory, and evaluation.
This diagram is a working model, not a deployed architecture.
observations
The system around the model
The question is whether coding and research agents get more reliable by treating their own output as something to execute, inspect, evaluate and revise, rather than returning the first plausible result.
The interesting work sits around the model rather than inside it: better verification, better retry decisions, execution feedback, memory worth keeping, search and branching, and lessons that survive a failure.
workflow
Attempt, verify, improve
- 01Task
A bounded goal with its constraints attached.
- 02Attempt
The agent works through tools, code, and search.
- 03Verify
Execution and tests decide, not a second opinion.
- 04Improve
A failure becomes feedback rather than a retry.
- 05Retain
Lessons worth keeping go to memory; the rest does not.
next questions
Open questions
- What feedback should survive between attempts, and what should not?
- What belongs in persistent memory rather than temporary task context?
- When is a deterministic verifier better than another model reviewing the work?
- How should an agent decide whether to retry, branch, search, or stop?
- How do you measure improvement without over-fitting to a weak evaluator?
- How much can improve without touching model weights?
- Where does human review belong in the loop?
observations
Current scope
The current scope is agent-system design around feedback, verification, memory, evaluation, retry decisions, and human review.
Still exploring
Thinking about the same problem?
I'm always interested in what works, what fails, and how to test the difference.