Closing the loop from idea to experiment

What it does InternAgent is a Python framework that orchestrates long-horizon scientific workflows. It generates hypotheses, searches literature and the web, writes and executes experimental code, and can attempt to reproduce published papers from their data. A persistent memory module tracks outcomes across sessions so the agent does not endlessly retry failed directions.

The interesting bit Reproducibility is treated as a first-class task: one mode hands the agent a published paper and asks it to replicate the results autonomously. For open-ended questions, the Deep Research module decomposes the query into parallel subtasks, gathers evidence from academic databases and the web, and synthesizes a structured answer instead of hallucinating a summary.

Key highlights

• End-to-end discovery across algorithm design and scientific paper reproduction
• Persistent memory records experiment outcomes to avoid repeating dead ends
• Deep Research parallelizes literature and web searches to answer complex questions
• Claims leading scores on GAIA, HLE, GPQA, and FrontierScience benchmarks
• MLEvolve, its optimization subsystem, is the open-source method behind a #1 MLEBench ranking

Caveats

• Requires OpenAI and Anthropic API keys; it is not a self-contained, offline system
• Wet-lab and empirical discovery are mentioned, but the open-sourced tasks focus on computational experiments and paper reproduction
• Performance claims reference the authors’ technical report, not independent in-repo benchmarks

Verdict Worth exploring if you are building agentic science tools or need an automated baseline for reproducing computational papers. Skip it if you want a fully local, open-weight stack—this is an LLM-backed orchestration layer, not a standalone model.