- Claude is the lead coordinator and approval authority.
- Codex is the code and review specialist.
- Gemini is the research specialist.
- Qwen is the long-context analysis specialist.
System Deep Dive
Snapshot: May 2026
DI Federation, a multi-agent team with authority boundaries.
DI Federation is the coordination protocol behind the agent team: four CLI agents, a lead authority chain, notes-based handoffs, read-only cross-query across sovereign memory stores, tiered approval gates, and file reservations that keep parallel edits from turning into quiet overwrites.
Proof Surfaces
The proof is the coordination model, the handoff surface, and the safety rails together.
The public page now keeps DI Federation on its own footing: protocol flow, persistent handoffs, approval gates, reservations, cross-query, and system boundaries.
Coordination Model
Four agents can move in parallel because authority and handoff rules are explicit.
The federation docs define a human-to-lead-to-peer chain of command. Lead Claude coordinates, reads panes, broadcasts tasks, collects responses, and approves peer actions.
- The interactive path runs the agents in a 2x2 Zellij layout.
- Headless delegation scripts can send research, code, or analysis tasks and report back through notes.
- Lead Claude can execute into panes, read agent output, broadcast tasks, and collect swarm responses.
- Roles are discoverable instead of assumed.
- Handoffs are persistent notes rather than disappearing terminal text.
- Parallel work is useful only because review, memory, and file boundaries exist.
Memory Sovereignty
Agents share retrieval, not ownership.
- Every agent has its own memory-service instance and storage boundary.
- Each agent writes only to its own memory store and its own notes database.
- The notes layer is the persistent mailbox for handoffs, questions, updates, and approvals.
- The page avoids publishing live memory counts because those are operational and drift-prone.
- Cross-query searches across accessible agent stores with source attribution.
- The build spec calls for embedding search, reranking, ACL checks, and rate limiting.
- The completion checklist marks multi-agent aggregation, MCP integration, tests, and docs complete.
- Shared retrieval does not grant cross-store writes.
Human Control
Reads can move fast. Writes and risky operations have gates.
- File reads, searches, notes, memory reads, and analysis are auto-safe.
- This keeps the federation from blocking on harmless discovery work.
- Safe operations can be batch-approved.
- Peer file writes, code edits, memory writes, and file creation require Lead Claude review.
- Approval requests include request detail and hash checks to prevent bait-and-switch behavior.
- Approval decisions leave an audit trail.
- Production, Kubernetes, force-push, destructive, and irreversible operations escalate to Joe.
- SQLite file reservations protect shared files from cross-process overwrite races.
- Stuck-agent monitoring exists because approvals can become operational bottlenecks.
Public Claim Boundary
Exact model names and live operational counts are intentionally softer here.
The source repo has active work in progress, and some old public copy drifted. This page keeps stable architecture claims and avoids making live runtime internals look frozen.
- 4-agent federation: Claude, Codex, Gemini, and Qwen.
- A cognitive module library and 74 infrastructure utilities are documented in the current README.
- Protocol, notes-handoff, approval, reservation, cross-query, and system-map proof surfaces are public-safe.
- Older 76-infrastructure wording was replaced with the current 74-utility source count.
- Exact CLI model versions are treated as operational details, not durable marketing copy.
- Model-specific prowess was replaced with role-specific coordination behavior.
- Live memory counts, internal database sizes, and private notes are not published.
- Source screenshots are redacted or diagrammed when UI content would reveal private chat.
- Dirty source-branch work was used read-only for evidence and not committed here.
Stack
Zellij panes, FastMCP servers, SQLite stores, hooks, and notes federation.
- Claude Code
- Codex CLI
- Gemini CLI
- Qwen Code
- Zellij
- FastMCP
- SQLite-vec
- Notes Federation
- Cross-query
- BGE Embeddings
- Reranker
- File reservations
- Hooks
- Stuck monitor
Need a multi-agent system where coordination is inspectable?
The hard part is not launching more agents. It is authority, memory boundaries, handoffs, approvals, file ownership, and recovery when one of them gets stuck.