# archimedes / studio
Autonomous research · Open source
Archimedes
Give it a prompt. It runs the research loop and hands you a trace you can verify.
- Three modes: orchestrated novelty, simple replication, and AlphaEvolve-style evolve
- A graph of specialised agents: ideation → novelty gate → plan → implement → check → reflect → write
- Novelty scored against the citation graph via arXiv and Semantic Scholar
input
Prompt
sparse MoE routing, low-resource
agent
Ideation
hypotheses
tool
Novelty gate
arXiv · Semantic Scholar
agent
Planner
staged experiments
tool
Implement
Claude Code · sandbox
gate
Success check
criteria vs metrics
agent
Reflector
next stage / revise
output
Writer
LaTeX manuscript
it keeps its failures
- Three modes: orchestrated novelty, simple replication, and AlphaEvolve-style evolve
- A graph of specialised agents: ideation → novelty gate → plan → implement → check → reflect → write
- Novelty scored against the citation graph via arXiv and Semantic Scholar
- 38 research replications published autonomously to the archimedes-run org
Design principle
Archimedes produces traces a human verifies, not finished papers. Transparency, steerability and trustworthiness come first. You give it a prompt and a mode; it runs a research loop end to end and leaves behind a literature map, an experiment plan, working code, metrics, the failures, and a manuscript draft.
Three modes
Orchestrated mode generates a novel idea, gates it against the citation graph, then plans and implements experiments stage by stage, reflecting on results before drafting a LaTeX manuscript. Simple mode hands a narrow task or strict paper replication directly to a Claude Code agent. Evolve mode is AlphaEvolve/FunSearch-style search: it samples a FAISS archive of past code variants, mutates the best one, and keeps whatever improves the empirical metric.
Architecture
Archimedes is a graph of specialised LLM agents — each with a narrow job, a strict prompt and a defined handoff — wired together to mirror how a research lab operates: ideation, novelty scoring, planning, Claude Code implementation, a success-criteria checker, a stage reflector and a writer. Long runs compress their event history so the loop can keep going for hours.
Proof of work
The archimedes-run GitHub organisation holds 38 repositories created by the system itself, including full replications of simulation-based inference and circle-packing experiments. I started archimedes.run as an open-source home for it, and keep building it in the open after hours.