Hive Mind Advanced

Hive Mind Advanced Skill

Queen-led hierarchical multi-agent coordination: a strategic queen coordinator directs specialized worker agents through collective decision-making (majority / weighted / Byzantine consensus) and a shared, persistent collective memory. This is the most structured coordination model in Claude Flow — reach for it when decisions need formal voting and agents need to share knowledge across a session.

Claude Code only: --claude below launches Claude Code with a coordination prompt, and MCP-flavoured framing throughout this skill assumes the Claude harness. On Codex / GPT-6 Astra: drop --claude and drive hive-mind spawn/ task/memory as plain CLI calls in one session; there is no claude-flow MCP wiring for Codex unless registered in ~/.codex/config.toml.

Quick path

Verified against the deployed CLI (claude-flow hive-mind --help / hive-mind <subcommand> --help, ruflo v3.38.21) — the --queen-type/--max-workers flags some upstream docs show do not exist on this build's spawn; topology, consensus, and agent cap are set at init, not per-spawn:

# 1. Initialize (topology + consensus + agent cap chosen here)
claude-flow hive-mind init -t hierarchical-mesh -c weighted -m 8

# 2. Spawn workers against an objective (pick a role, optionally launch Claude Code)
claude-flow hive-mind spawn -n 8 -r worker --claude -o "Build microservices architecture"

# 3. Monitor (there is no `hive-mind metrics` subcommand — use status + memory)
claude-flow hive-mind status
claude-flow hive-mind memory

Consensus (init -c): majority, weighted (queen 3x), byzantine (2/3 supermajority, the default). The "queen type" framing (strategic for research/planning, tactical for implementation, adaptive for optimisation) is conceptual in this CLI build — it shapes the objective/prompt you pass, not a flag. Memory persists to RuVector PostgreSQL (pgvector/HNSW) in production, SQLite locally.

When to use

Use the Hive Mind when the work genuinely needs:

• Formal consensus on critical decisions (architecture choice, tech stack, release readiness) — not just parallel task execution.
• Collective memory shared across agents and resumable sessions.
• Queen-led hierarchy — a coordinator that decomposes objectives and assigns specialized workers (researcher, coder, analyst, tester, architect, reviewer, optimizer, documenter).

When not to use

• Topology-based agent coordination (mesh, hierarchical, star, ring) without formal voting → swarm-advanced is simpler.
• A structured development lifecycle (spec, architecture, TDD, review, deploy) → sparc-methodology (17 development modes).
• Full development + quality-engineering pipelines → build-with-quality.
• Cloud-based swarm deployment / workflow automation → flow-nexus-swarm.
• GitHub-specific PR review coordination → github-code-review.

Reference material

Detailed guidance lives in references/ — load the file for the task at hand:

• references/architecture.md — architecture patterns, worker specialization, collective memory system, RuVector PostgreSQL backend, consensus mechanisms.
• references/operations.md — init / spawn / monitor commands, session management, consensus building, collective-memory operations, task distribution, auto-scaling, and integration with Claude Code / SPARC / GitHub.
• references/performance-and-config.md — memory & database optimisation, benchmarks, hive/memory config schemas, hooks integration, and best practices per queen type.
• references/api-and-cookbook.md — troubleshooting (memory / performance / consensus), advanced topics (custom workers, neural training, multi-hive, export/import), the HiveMindCore / CollectiveMemory / HiveMindSessionManager API reference, worked examples, and skill progression.

Related skills

• swarm-advanced — general swarm coordination (no consensus required)
• consensus-mechanisms — distributed decision making
• memory-systems — advanced memory management
• sparc-methodology — structured development workflow
• github-code-review — repository PR review and coordination

Skill Version: 1.0.0 Last Updated: 2026-07-28 Maintained By: Claude Flow Team License: MIT