- Claude (Anthropic Opus 4.6) — Lead coordinator with full authority
- Codex (OpenAI GPT-5.3) — Code specialist, requires approval
- Gemini (Google Gemini 3 Flash) — Research specialist, read-only
- Qwen (Alibaba Qwen3-Coder-480B) — Deep analysis, requires approval
System Deep Dive
Snapshot: March 2026
DI Federation, four AI agents with sovereign memory and constitutional control.
A multi-agent orchestration system where Claude, Codex, Gemini, and Qwen each run in their own terminal pane with sovereign persistent memory, communicating through a notes-based federation protocol. Hub-and-spoke authority, constitutional identity locks, a Raw Reality Index for failures, and a 46-module cognitive architecture that gives the lead agent metacognition, theory of mind, and scar-based learning.
The Federation
Four labs. Four agents. One coordination layer.
Each agent writes only to its own sovereign memory store. Cross-agent access is read-only. Authority flows through the lead, not through consensus.
- Notes-based inter-agent messaging with 60 federation tools
- Cross-query bridge with BM25 + vector hybrid search (19 tools)
- Hub-and-spoke authority — only the lead can approve peer actions
- Ephemeral workers (claude-2 through claude-5) spawned on demand
Cognitive Architecture
Metacognition, scars, theory of mind, and identity.
- Metacognition and self-model for reasoning about reasoning
- Scar registry — permanent learning from mistakes with decay
- Narrative identity for coherent selfhood across sessions
- Epistemic boundary guard — knowing what it does not know
- Theory of mind — modeling Joe's mental state and intent
- Preference learning — adapting to patterns over time
- Emotional momentum — tracking affective context
- Collaboration pattern memory across sessions
- Surprise quantification — measuring unexpected outcomes
- Confidence tracker — honest uncertainty quantification
- Counterfactual logging — tracking "what if" paths not taken
- Growth tracking — measuring improvement over time
Constitutional Layer
Some things cannot be changed by consensus.
- Identity, scars, goals, Joe's relationship, safety rules, and human authority can never be modified by federation consensus
- Coalition prevention — no peer group can override the human
- Epistemic firewalls — agents cannot alter each other's beliefs
- Operating modes: SEAL (locked), EVOLUTION (sandbox), MAINTENANCE (emergency)
- Phoenix Protocol — graceful crash recovery with epoch-based message validation
- Raw Reality Index — stores failures, dead ends, and contradictions explicitly
- Dream mode — biological sleep-inspired background consolidation every 4 hours
- 76 infrastructure modules: retry, backoff, circuit breakers, failover, dedupe
Surfaces
Dashboard, MCP servers, and a full skills arsenal.
- Timeline, Search, Federation, 3D Graph, Control Room
- Dream Mode, Team Chat, Notes, Tasks, Approvals
- Health, Traces, Metacognition, Threat Board, Directives
- Notes-federation (60 tools) and cross-query bridge (19 tools)
- Embedding gateway (BGE-large-en-v1.5 on FastAPI)
- Dream mode, memory provenance, semantic dedup, tracing
- 4 sovereign SQLite-vec stores (ports 8100–8103)
- Cross-query bridge for read-only federation search
- Canonical facts index for deduplication
- 7,500+ memories, 102MB total across all agents
Stack
Built from hooks, signals, and federation primitives.
- Claude Code
- Codex CLI
- Gemini CLI
- Qwen Code
- Zellij
- SQLite-vec
- FastMCP
- Node.js
- Python
- React
- Vite
- Ollama
- BGE Embeddings
- Kuzu Graph
Need a multi-agent system with real authority controls?
That means constitutional design, not just a prompt chain.