Agents Overview

AutoResearch uses six specialized agents plus an orchestrator. Each agent has a specific role, runs on a specific LLM, and operates within strict boundaries.

At a Glance

Agent	LLM	Type	One-Line Role
Orchestrator	Claude Opus	Persistent session	Decides what to do, dispatches work
Planner	Claude Opus	Sub-agent	Designs experiments, decomposes tasks
Writer	Claude Opus	Clean session	Writes paper sections in LaTeX
Scout	Gemini	omc team 1:gemini:scout	Searches literature, generates ideas
Coder	Codex	omc team 1:codex:coder	Implements code, runs experiments
Judge	Codex	codex exec -m gpt-5.4	Evaluates artifacts, renders verdicts

graph TB
    O[Orchestrator<br/>Claude Opus]
    
    O --> P[Planner<br/>Claude Opus]
    O --> W[Writer<br/>Claude Opus]
    O --> SC[Scout<br/>Gemini]
    O --> C[Coder<br/>Codex]
    O --> J[Judge<br/>Codex]

    P -.->|"plan.md"| C
    SC -.->|"papers.bib"| W
    C -.->|"results.yaml"| J
    J -.->|"verdict.yaml"| O

    style O fill:#f9f0ff,stroke:#7c3aed,stroke-width:2px
    style P fill:#ede9fe,stroke:#7c3aed
    style W fill:#ede9fe,stroke:#7c3aed
    style SC fill:#ecfdf5,stroke:#059669
    style C fill:#fef3c7,stroke:#d97706
    style J fill:#fef3c7,stroke:#d97706

Solid arrows = dispatch commands. Dotted arrows = data flow via disk.

Agents never talk to each other directly. The Orchestrator dispatches tasks, and agents exchange data through files in .omc/research/.

Agent Interaction Matrix

Who sends work to whom, and through what mechanism:

From \ To	Orchestrator	Planner	Writer	Scout	Coder	Judge
Orchestrator	—	dispatch	dispatch	dispatch	dispatch	dispatch
Planner	result	—	—	—	—	—
Writer	result	—	—	—	—	—
Scout	result	—	—	—	—	—
Coder	result	—	—	—	—	—
Judge	verdict	—	—	—	—	—

All arrows point to/from the Orchestrator. No agent dispatches work to another agent. No agent reads another agent's context.

Key Interaction Chains

Ideation Chain

Orchestrator → Scout (find related work)
            → Orchestrator (synthesize idea)
            → Judge (evaluate idea)
            → Orchestrator (decide: refine or advance)

Implementation Chain

Orchestrator → Planner (design experiment)
            → Coder (implement design)
            → Orchestrator (review code)
            → Judge (evaluate code quality)
            → Orchestrator (decide: fix or advance)

Writing Chain

Orchestrator → Scout (fetch paper details)
            → Writer (draft section)
            → Judge (review draft — three-model panel)
            → Orchestrator (decide: revise or advance)

Error Recovery Chain (ralph)

Coder encounters error
→ Coder attempts self-fix (up to 3 tries)
→ If still failing → Orchestrator intervenes
→ Orchestrator may re-dispatch with different instructions
→ Or escalate to human

The Orchestrator is always in the loop

Every chain starts and ends at the Orchestrator. This ensures centralized decision-making and a complete audit trail in orchestrator.log.

Detailed Agent Pages

• Orchestrator — the brain
• Planner — experiment design
• Writer — paper writing
• Scout — literature and ideas
• Coder — implementation
• Judge — evaluation and review