Agent self-testing adds verification steps to an agent's workflow. After producing an artifact or changing an environment, the agent can run the result, inspect visible behavior, compare it with requirements, and use discovered defects to guide revisions.
Self-testing improves the evidence available to the agent, but it is not independent evaluation. The same system may misunderstand both the task and the test, overlook failures, or accept weak criteria, so important results still benefit from external review.
ELI5
Agent self-testing means an AI helper checks the thing it made instead of assuming the first attempt is correct. It uses the result of the check to decide whether more work is needed.
For example, after creating a game movement control, the helper could run the game, try the control, notice that it moves in the wrong direction, and revise the settings.
