Skip to content

Latest commit

 

History

11 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

Hive v4.0: Bio-Inspired Swarm Orchestrator for Claude Code

A multi-agent orchestration skill that coordinates parallel Claude Code agents using mechanisms from ant colonies, honeybee swarms, and AI research.

What It Does

/hive <task> breaks your task into subtasks, spawns agents in parallel waves, coordinates them through shared findings, resolves conflicts, and learns from each run.

Skills in This Plugin

Skill What it does
/hive Parallel swarm orchestration for execution tasks: research, fixes, refactors, audits.
/hiverev Persona-driven review of any artifact from any angle: a UI, a landing page, a cold email, a README, a pitch deck, a flow. Pick who is looking and what lens they bring, get a prioritized friction punch list. Runs in-context, no sub-agents.
/hivesim Strategic wargame engine. 7 phases: scout, decompose, quantitative buyer sim (10-15 personas), deep buying journeys, 3-adversary red team with FATAL verification, pre-mortem, and a one-page Go/No-Go Decision Brief with kill criteria.

What's New in v4.0 (April 8, 2026)

  • Concurrency unlocked: Old limit was 2 heavy agents (from a March 2026 incident). Stress-tested to 25 read / 20 write / 15 mixed agents with zero failures across 161 launches. Default start concurrency raised from 5 to 10.
  • Mechanism audit: Honest assessment of all 27 mechanisms. 10 fire regularly, 17 are theoretical/broken. Dead mechanisms documented, not removed (platform may support them in future).
  • Cross-skill sync: Updated concurrency limits across /hivesim, /sim, /gtm, /hivefront. All skills now use proven ceilings.
  • Hivesim personas: Now spawn 10-15 personas per wave instead of 2. ~4x faster wall clock time for strategic analysis.

Performance (April 2026 benchmarks)

Scenario Agents Wall Time Speedup vs Sequential
10-file audit 10 haiku ~60s 8x
15-fix code wave 2 opus + 4 sonnet + 9 haiku ~200s 4.4x
109-file catalog 5 waves x 25 haiku ~3 min 18x
20-file write batch 20 haiku ~73s 10x

Proven Concurrency (April 8, 2026 test suite: 78 agents, 10 experiments, 0 errors)

Config Agents Result Avg Latency
5 haiku 5 PASS 5.4s
10 haiku 10 PASS 9.2s
4 sonnet 4 PASS 9.7s
2 opus 2 PASS 11.9s
10h + 3s 13 PASS h:7.6s / s:12.9s
8h + 1o 9 PASS h:5.5s / o:10.5s
5h + 2s + 1o 8 PASS h:4.9s / s:7.1s / o:10.7s

Key findings:

  • No hard 529 wall found up to 13 mixed agents
  • Model tiers don't conflict (haiku/sonnet/opus share the same pool)
  • 10+ agents get queued ~2 min before executing (Claude Code internal queue)
  • 529 errors are transient API load spikes. Always retry with 30s delay.
  • Safe default: 13 mixed agents. Flat count, not weighted.

Installation

In Claude Code:

/plugin marketplace add CipherandRow/claude-hive-plugin
/plugin install hive@cipherandrow

Or manually copy any skill folder from skills/ into your ~/.claude/skills/ directory.

Usage

/hive research the top 5 competitors and compare pricing
/hive fix all failing tests in src/
/hive --dry-run refactor the auth module
/hive --resume
/hive --isolate update all API endpoints to v2

Flags

Flag Description
--resume Resume from last checkpoint
--isolate Force git worktree isolation for all agents
--dry-run Show execution plan without launching agents
--verbose Show mechanism trace after each wave (default: on)
--quiet Suppress execution trace

How It Works

  1. Mode Detection: Counts subtasks, selects Lite/Standard/Full mode
  2. Strategy Selection: Reads past run history, picks best strategy
  3. Planning: Breaks task into waves with dependency ordering and reserve pool
  4. Pre-flight: Context ceiling check, confidence calibration, agent TTL calculation
  5. Execution: Spawns agents in waves with stigmergy (shared findings board), velocity scaling, and error handling
  6. Synthesis: Merges results using decision protocols (vote, consensus, or AAD)
  7. Learning: Appends scored results to history, updates playbook with time-decay

Mechanisms

Active (fire regularly, proven in production)

# Mechanism Origin Purpose
2 Self-Validation Gates Leaf-cutter ants Agents validate own output before returning
3 Reasoning Tree Conflicts AgentAuditor Find exact divergence point in reasoning
4 Stigmergy Ants Shared findings for indirect coordination
6 Semantic Quorum Honeybees + LLM Early commitment via semantic similarity
7 Scout Retirement (TTL) Honeybees Kill stuck agents proactively
11 Cross-Inhibition Honeybee stop signals Competing proposals dampen each other
13 Assembly Line QC Leaf-cutter ants Pipeline handoff quality gates
15 Adaptive Mode Ant response thresholds Auto-scale mechanisms to task size
17 Auto-Verify Gate Manufacturing QC Run tsc + tests after code-writing waves
27 Debug Trace Distributed tracing JSONL event log for every decision

Theoretical (defined but rarely/never fire in current Claude Code)

# Mechanism Why Inactive
1 Pheromone Evaporation History is read but scoring never changes agent routing
5 Completion Velocity Concurrency is now fixed at proven tiers, not dynamically scaled
8 Decision Protocols Vote/consensus/AAD never explicitly activates as distinct step
9 Swarm Playbook No evidence of playbook influencing future runs
10 Ready-Up Signal Pre-flight exists but not as a distinct mechanism
12 Inspector Agents Never fires in production
14 Checkpoint/Resume Defined but never used across 50 runs
16 Worktree Isolation Broken on Windows (file delivery fails on cleanup)
18-26 Advanced mechanisms Platform doesn't yet support (JSON schema enforcement, dynamic tool loading, hot reassignment, etc.)

These remain in the skill definition for forward compatibility. As Claude Code and the Agent SDK evolve, more mechanisms may activate.

Model Selection (data-driven)

Model Avg Duration Avg Tool Calls Best For
Haiku 7-29s 1-5 File reads, grep, single-file fixes, summaries
Sonnet 85-95s 20-25 Multi-file fixes, code review, medium complexity
Opus 94-115s 10-12 Complex rewrites, architecture decisions, synthesis

Prompt Discipline (proven by A/B test)

Strict prompts beat loose prompts in every dimension:

  • 44% faster per-agent execution
  • 10x less code bloat
  • 0 scope violations vs 1

Always include in write agent prompts:

ONLY EDIT THIS FILE: [path]. DO NOT edit any other file.
DO NOT add inline comments, refactor, or change code outside the task.

Requirements

  • Claude Code CLI
  • Git (for worktree isolation, optional)

Version History

Version Date Changes
v4.2 July 23, 2026 Added /hiverev (persona-driven review of any artifact) and /hivesim (strategic wargame engine) skills
v4.1 April 9, 2026 Full concurrency test suite (78 agents, 0 errors), flat ceiling model replaces weighted tokens
v4.0 April 8, 2026 Concurrency unlocked (2->25), mechanism audit, cross-skill sync
v3.0 March 25, 2026 14 mechanisms, semantic quorum, reasoning tree, A2A protocol
v2.0 March 2026 Adaptive mode, worktree isolation, stigmergy
v1.0 March 2026 Basic parallel agent orchestration

License

MIT

About

Claude Code plugin with three skills: /hive (bio-inspired swarm orchestrator, 16 mechanisms from ant colonies and honeybee swarms), /hiverev (persona-driven review of any artifact), /hivesim (strategic wargame engine with red team and pre-mortem).

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors