Overview
The loop begins with a baseline. The system is measured before any changes are made. The agent then proposes a small hypothesis, such as changing one implementation detail to improve speed. It modifies only the allowed part of the system and runs the assessment again.
If the measured result improves under the defined constraints, the change is retained. If it worsens or fails, the change is reverted. In both cases, the agent records what it tried and what happened in a research log. The next iteration uses that accumulated history to avoid repeating failed ideas and to build on successful ones.
The loop is:
- Read the current state and the experiment log.
- Propose one bounded change.
- Apply the change.
- Run the assessment.
- Compare the new result with the baseline or current best.
- Keep or revert the change.
- Record the hypothesis, modification, score, and outcome.
- Select the next hypothesis.
The assessment must represent the actual objective closely enough to guide the loop. In a website-speed example, the agent can modify the site, run a standardized performance test, and retain changes that improve the measured loading metrics. Additional constraints can prevent the agent from winning the metric by damaging the product, such as requiring the visual result to remain unchanged.
Iteration speed matters. A change that takes seconds and an assessment that takes seconds allow many experiments. A change that takes an hour followed by an hour-long assessment can still be automated, but the improvement cycle will be much slower.