Scientify
Open workspace
Back to home

AI Scientist research loop

The agent advances a project the way research works

Research methods have unknown effects and usually fail. Built on Codex, the scientific agent explores several hypotheses at once, shares findings across branches, and runs adversarial review on every conclusion as it iterates on one project.
Start researching

Research loop

Literature → Hypotheses → Parallel experiments → Review → Iterate

The core agent is open source with 2k+ GitHub stars; the related paper was accepted at ICML 2026.

How it works

01

From literature to testable hypotheses

The agent plans search strategies around the project, tries to retrieve full-text PDFs, extracts testable hypotheses and writes an experiment plan.

  • Search and full text
  • Testable hypotheses
  • Experiment plan
02

Several hypotheses explored in parallel

Several subagents test different hypotheses at the same time. A problem found on one branch, such as a data defect, is shared with the others so they do not repeat it.

  • Parallel subagents
  • Communication across branches
  • Shared failure lessons
03

Review before the next round

Every conclusion passes adversarial review before the next iteration. Literature, code, data and results stay in one workspace.

  • Adversarial review
  • Continuous iteration
  • Reproducible results

Start from a research goal

Start researching
Explore next