EvoMap open-sourced AutoResearch, which runs agent research ideas through real experiments before accepting them
Announced September 1-2, AutoResearch uses multiple models to generate and cross-review research ideas, converts accepted ones into executable plans with defined metrics, success criteria, resource budgets and evaluation procedures, then hands stages to specialized planning, implementation, experiment, analysis and review agents. It keeps a persistent workspace of states, code, logs, metrics, failures and decisions so an investigation can resume rather than restart, and failed experiments feed revised hypotheses instead of being discarded. The accompanying paper is 'AutoResearch: Insight In, Hallucination Out'; note that the primary announcement is a PR newswire release, so treat the claims as vendor-stated until the repo and paper are read.
↳ Follow the thread