Agent Architecture

The ULTRATHINK research system uses a hierarchical Claude Code agent architecture with specialized agents per project and wave coordinators.

Agent Hierarchy

orchestrator-agent (Opus) ─── Master coordinator
    |
    |── Core Agents (Sonnet)
    |   |── rosa-agent ──────────── Papers/paper/
    |   |── grandmother-agent ───── Papers/paper3/
    |   |── spikegpt-agent ──────── Papers/paper4/
    |   |── memory-agent ────────── Papers/paper6/
    |   |── imagination-agent ───── Papers/paper7/
    |   └── audio-agent ─────────── Papers/paper8/
    |
    |── Wave 2 Agents (paper26-35, one per project)
    |   |── spikemoe-agent, neuromamba-agent, tempobert-agent, ...
    |
    |── Wave 3 Agents (paper36-45, one per project)
    |   |── spikessl-agent, neurogan-agent, spikediff-agent, ...
    |
    |── Wave Coordinators
    |   |── audio-snn-agent ────── Wave 4 (paper46-60)
    |   |── thorpe-tb-agent ────── Wave 5 (paper61-80)
    |   |── tree-snn-agent ─────── Wave 6 (paper81-95)
    |   |── deepseek-snn-agent ─── Wave 7 (paper96-110)
    |   └── domain-snn-agent ───── Wave 8 (paper111-130)

Model Tiers

Agent Configuration

Agents are defined in ~/.claude/agents/. Each agent has:

• Project-specific context and instructions
• Access to the project's directory
• Knowledge of the review framework

Multi-LLM Verification

For significant changes (>50 lines), the multi-llm-verify skill runs consensus checks across Claude, Codex, and Gemini. Results are logged to ~/.claude/verification-logs/.