Kosmos — автономный ИИ-агент для научных исследований

★ 7.0 · vibe-coding

Kosmos is an open-source autonomous AI scientist that implements the architecture described in Lu et al. (2024) for end-to-end scientific discovery. It runs multi-cycle research workflows: searching literature across ArXiv, PubMed, and Semantic Scholar, generating testable hypotheses, designing experiments, executing AI-written code inside sandboxed Docker containers, and validating findings with an 8-dimension ScholarEval quality framework. The system coordinates six specialized agents — Research Director, Hypothesis Generator, Experiment Designer, Data Analyst, Literature Analyzer, and Plan Creator — alongside an optional Neo4j knowledge graph and a context compression pipeline achieving roughly 20:1 ratios. Over 100 LLM providers are supported via Anthropic, OpenAI, and LiteLLM, with built-in budget enforcement through a configurable BUDGET_LIMIT_USD cap. Kosmos is aimed at researchers and data scientists looking to automate hypothesis generation, literature synthesis, and experimental validation in domains like biology, materials science, or chemistry.