Overview
AutoResearch is a multi-agent orchestration system that automates deep learning research from initial idea through to a finished paper. It coordinates three large language models — Claude Opus, Codex (GPT), and Gemini — through a structured seven-stage research pipeline.
Core Principle
The Orchestrator decides. Agents execute.
AutoResearch follows a strict separation between decision-making and execution. The Orchestrator (Claude Opus, running in the main Claude Code session) holds the research vision and decides what to do next. Specialized agents handle how to do it, each operating within clear boundaries.
No agent acts on its own initiative. Every agent invocation is triggered by the Orchestrator, and every result is reported back to the Orchestrator for the next decision.
Three LLMs, Three Roles
| LLM | Strengths | Primary Roles |
|---|---|---|
| Claude Opus | Deep reasoning, long context, nuanced judgment | Orchestrator, Planner, Writer |
| Codex (GPT) | Fast code generation, structured evaluation | Coder, Judge |
| Gemini | Broad knowledge, literature access | Scout (literature search, idea generation) |
Why three models?
No single model excels at everything. Claude reasons deeply but writes code slowly. Codex generates code fast but lacks research judgment. Gemini has broad knowledge but needs structured tasks. AutoResearch plays to each model's strengths.
Built on OMCC
AutoResearch runs on top of the OMCC (Orchestrated Multi-Claude Coordination) harness. OMCC provides:
- Tmux-based agent sessions — persistent, named sessions for each agent
- Heartbeat monitoring — the Orchestrator knows if an agent is alive
- State-on-disk conventions —
.omc/directory structure for persistent state - Hook system — pre/post hooks for quality gates and enforcement
- Cross-session communication — agents exchange data through disk, not memory
Key Features
- Seven-stage pipeline — Ideation, Design, Implementation, Training, Analysis, Writing, Review
- Configurable automation gates — human approval, auto-judge, or fully automatic per stage
- Three orthogonal execution modes — autopilot, ralph, ultrawork
- Two-tier information architecture — context for working memory, disk for long-term memory
- Cross-LLM review — every artifact is reviewed by a different LLM than the one that created it
- Two-phase training monitoring — active watch during start, periodic patrol after stabilization
- Session-resilient state — crash, restart, resume without loss
- Workspace isolation — multiple research projects in parallel
What you need
AutoResearch requires Claude Code (with Claude Opus), Codex CLI, and Gemini CLI installed and configured. See Quick Start for setup.
Next Steps
- Architecture — understand the system topology
- Quick Start — run your first research session
- Agents — learn about each specialized agent
- Pipeline — explore the seven research stages