Autoresearch is an agent-driven improvement loop built around executable feedback. A human or system defines the target, constraints, and evaluator; an agent proposes a change, runs the experiment, records the result, and keeps or discards the change based on measured evidence.

In AI-assisted software development, the pattern moves research automation into ordinary code work: performance tuning, benchmark improvement, bug hunts, prompt optimization, repository simplification, and gated migrations. The useful boundary is not full autonomy, but whether the loop has a trustworthy evaluator and enough trace evidence for humans to audit the search.

In the News