Skip to content
D-CSIL

Project File

Multi-Agent Systems

Councils, hives, and debate: multiple specialized models coordinating on one problem — from a Raspberry Pi MQTT hive to a five-model local debate framework.

Active ResearchStructure as a substitute for scale.

CSI Agent Hive

OBSERVED

A distributed system where Raspberry Pi 5 nodes host agents with distinct personalities (logical, emotional) communicating over MQTT. Agents react probabilistically to peers and maintain per-pair trust scores updated by sentiment analysis — trust as an emergent, measured quantity.

Multi-Agent Chat / Debate UI

OBSERVED

A Streamlit framework running five local Ollama models (Qwen2, Llama3, Mistral, Gemma, DeepSeek) in turn-based debate with configurable personas, an orchestrator with intervention powers, RAG context injection, and a complete bulk HR resume-review workflow (up to 1,000 documents with translation).

SuperRats

EXPERIMENTAL

A vendor-neutral capability-based execution framework: route to the cheapest capable model first, run independent sub-agents concurrently, escalate on low confidence, verify code before accepting it. Its experience layer now feeds Remy's hive via an ExistingRemyBackend.

The Sub-Agent Hive

PRINCIPAL INVESTIGATOR

human in the loop

REMY · MODEL ROUTER

spawn_agents · routing · fallback chains · budgets

POLICY ENGINE + BUDGETS

deterministic allow/deny · $0.02 cloud ceiling enforced live

RESULT → CORTEX WRITEBACK

Hover or tap an agent to see its responsibilities.

Current Limitations

  • No controlled council-vs-monolith benchmark has been run yet — the comparison is designed, not measured.
  • Hive trust dynamics are observed qualitatively; no formal metrics published.

Next Steps

  • +Fold SuperRats routing fully into Remy's spawn_agents.
  • +Design a measurable council-vs-single-model evaluation.