Automated experiment execution converts a research plan into repeatable actions such as generating code, setting parameters, launching runs and checking completion. It reduces manual handling and makes it practical to test more configurations.
Automation must preserve the exact setup and output needed for reproducibility. A fast run is not useful when errors are hidden, conditions drift or the recorded result cannot be traced to the experiment that produced it.
ELI5
Automated experiment execution lets software carry out a defined experiment with less manual work. It can prepare settings, launch runs, watch progress, and save the results in a repeatable way.
For example, an agent can test ten learning rates for a model and record the result of each run. The automation must preserve the exact settings and errors so another researcher can understand and repeat what happened.
What should an automated experiment record?
It should record code or procedure versions, parameters, environment, inputs, timestamps, outputs, errors and the checks used to determine completion.
Can an agent change an experiment while it is running?
It can adapt within approved rules, but changes should be logged and separated from the original condition so results remain interpretable.
