Let AI be a pilot in science: five guardrails for responsible discovery

AI agents can accelerate scientific work but should act as pilots rather than full automation. SciSciGPT demonstrates a multi‑agent workflow (ResearchManager, literature review, data analysis, planning) with an EvaluationSpecialist auditing outputs and provenance logging, delivering faster, higher‑quality results while preserving human judgment. The piece argues for collaboration over automation, a new ethics of transparency and accountability, field‑specific standards, cross‑disciplinary training, and trust engineering to keep science credible, interpretable and reproducible.