Software-agent research is getting stricter about signal, evidence, and risk
The period’s strongest theme is tighter control over software agents: cleaner traces for training, better logs for evaluation, and harder tests for what agents do in real repositories and real workspaces.