AI research agents keep shipping work that hasn't been checked against reality.
AutoResearch is an open-source framework that tries to close that gap. It couples sandboxed Python and PyTorch execution with iterative code repair, so generated experiments are actually run and fixed before they're considered done. On the citation side, it verifies that referenced sources support the claims made — something most current agents skip entirely. The system treats runtime errors, failed citation checks, and reviewer feedback as filters, not obstacles. In controlled evaluations across HumanEval, MBPP, a SciCode subset, and citation-validation tasks, AutoResearch outperformed directly comparable baselines on execution success, citation validity, and workflow completion.
The reliability problem it targets is real: AI agents generating plausible-looking but unverified research artifacts have become a quiet credibility crisis across labs and academia. A framework that enforces execution grounding and claim auditing before output is finalized could meaningfully raise the floor on what automated systems produce — and give researchers a defensible audit trail.
The authors are careful to frame AutoResearch as a reliability-oriented assistant, not a fully autonomous scientist — a distinction worth noting given how many similar projects oversell the autonomy angle before the hard problems are solved.