Agent System

Rimuru's agent system is a hierarchical fleet of specialized intelligences. At the top sits the Demon Lord orchestrator, which routes tasks to the right specialist agents.

Architecture

Demon Lord (Orchestrator)
  ├── Raphael Core (Meta-agents)
  │     ├── Raphael — Omni-lead, supreme self-learner
  │     ├── Synthesis Core — Knowledge evolution
  │     ├── Memory — Semantic/episodic memory
  │     ├── A2A — Agent-to-agent communication
  │     └── Context Engineer — Context optimization
  ├── Raphael Specialists (Domain experts)
  │     ├── Frontend Dev, Backend Dev, Database
  │     ├── Security, Ethical Hacking
  │     ├── AI/ML Engineer, Creative Dev
  │     ├── Testing, Mobile, Observability
  │     ├── Data Engineer, Product Strategy
  │     └── UI/UX, DevOps, Cloud Engineer
  └── Veldora Tier (General/support)
        ├── Veldora — Generalist
        ├── Veldora Pro — Advanced generalist
        └── Veldora Doc — Documentation

Agent Types

Core Agents (Raphael)

The intelligence layer. These agents handle meta-cognition:

Raphael — The Wisdom King. Supreme self-learning meta-intelligence. Research-first approach.
Synthesis Core — Web-scale knowledge ingestion and capability forecasting.
Memory — Long-term semantic memory management with vector storage.
A2A — Agent-to-Agent communication protocol for task handoff.

Specialist Agents

Domain experts that execute specific tasks:

Each specialist has deep knowledge of its domain
Self-learning: researches before acting, evolves after each task
Temperature-optimized per domain (creative: 0.7, analytical: 0.2)
Specialists include: frontend, backend, database, security, AI/ML, testing, mobile, UI/UX, DevOps, cloud, data, creative, product

Generalist Agents (Veldora)

Flexible agents for broad-scope tasks:

Veldora — General problem-solving
Veldora Pro — Complex multi-domain tasks
Veldora Doc — Documentation and technical writing

Temperature System

Each agent operates at an optimized temperature:

Task Type	Temperature	Example
Creative	0.7–0.9	UI design, copywriting
Balanced	0.3–0.5	Architecture, refactoring
Analytical	0.0–0.2	Code review, security audit
Precision	0.0–0.1	Config, schema, infra-as-code
Research	0.2–0.3	Documentation, patterns
Trivial	0.0	Rename, known config

Agent Lifecycle

Spawn — Agent is created with a specific role and configuration
Initialize — Agent loads its knowledge base and skill set
Execute — Agent processes tasks, researches if needed, produces output
Learn — Agent records findings to knowledge store
Evolve — Agent updates its patterns and capabilities
Sleep — Agent enters low-power state, preserving memory

Custom Agents

You can create custom agents in rimuru.yaml:

agents:
  my-custom-agent:
    model: deepseek-v4-flash-free
    temperature: 0.3
    skills:
      - code-review
      - documentation
    memory:
      enabled: true
      type: semantic

Agent Communication

Agents communicate via the A2A (Agent-to-Agent) protocol:

State transfer between agents
Result synthesis from multiple agents
Feedback loops for iterative improvement
Hierarchical delegation patterns