Pop Power Mode

Multi-agent orchestration using Claude Code's native background agents for true parallel collaboration.

Core principle: Agents work in parallel via run_in_background: true, share discoveries via file, and coordinate through sync barriers.

Architecture: Native Async (Claude Code 2.0.64+)

┌─────────────────────────────────────────────────────────────────┐
│                   NATIVE ASYNC POWER MODE                        │
├─────────────────────────────────────────────────────────────────┤
│   ┌───────────┐  ┌───────────┐  ┌───────────┐                  │
│   │  Agent 1  │  │  Agent 2  │  │  Agent 3  │                  │
│   │background │  │background │  │background │                  │
│   └─────┬─────┘  └─────┬─────┘  └─────┬─────┘                  │
│         └──────────────┼──────────────┘                         │
│                        │                                        │
│                ┌───────▼───────┐                               │
│                │  Main Agent   │  ← Coordinator (TaskOutput)   │
│                │  (Polling)    │                               │
│                └───────────────┘                               │
├─────────────────────────────────────────────────────────────────┤
│ Requirements: Claude Code 2.0.64+ (no external dependencies)   │
│ Setup: Zero config - just run /popkit:power start              │
└─────────────────────────────────────────────────────────────────┘

Tier Comparison

Feature	Free	Premium ($9/mo)	Pro ($29/mo)
Mode	File-based	Native Async	Native Async
Max Agents	2	5	10
Parallel Execution	Sequential	✅ True parallel	✅ True parallel
Sync Barriers	Basic	✅ Phase-aware	✅ Phase-aware
Insight Sharing	Basic	✅ Full	✅ Full
Redis Fallback	❌	❌	✅ Optional

Free Tier: File-Based Mode

Free tier users get file-based coordination:

• Works with 2 agents (sequential)
• Shared context via .claude/popkit/insights.json
• Good for learning Power Mode concepts
• Zero setup required

Premium Tier: Native Async Mode

Premium users unlock native async capabilities:

• Up to 5 agents working in true parallel
• Uses Claude Code's Task(run_in_background: true) API
• Main agent polls via TaskOutput(block: false)
• Phase-aware sync barriers
• Full insight sharing

Pro Tier: Full Power

Pro users get maximum capabilities:

• Up to 10 parallel agents
• Optional Redis fallback for high-volume scenarios
• Team coordination features

Overview

Inspired by:

• ZigBee/Z-Wave mesh networks (failover, redundancy)
• DeepMind's objective-driven agents (constrained exploration)
• Node-RED (flow-based coordination)

When to use:

• Complex tasks requiring multiple specialized agents
• Tasks that benefit from parallel exploration
• When agents need to share discoveries in real-time
• Large implementations with distinct subtasks

When NOT to use:

• Simple, single-agent tasks
• Tasks requiring tight sequential dependency
• When Redis is unavailable
• Quick fixes or small changes

Prerequisites

Native Async Mode (Default - Zero Config)

• Claude Code 2.0.64 or later
• Premium or Pro tier
• No additional setup required!

Activation

Method 1: Command (Recommended)

/popkit:power start "Build user authentication with tests and docs"

Method 2: Issue-Driven

/popkit:dev work #45 -p

Method 3: Auto-Detection

# Power Mode auto-enables for issues labeled "epic" or "power-mode"
/popkit:dev work #45

How It Works

1. Define Objective

Power mode requires a clear objective with:

• Description: What we're building
• Success criteria: How we know we're done
• Phases: Ordered stages of work
• Boundaries: What agents can't do

Objective:
  description: "Build user authentication"
  success_criteria:
    - "Login endpoint works"
    - "Tests pass"
    - "Documentation updated"
  phases: [explore, design, implement, test, document]
  boundaries:
    - file_patterns: ["src/auth/**", "tests/auth/**"]
    - restricted_tools: []

2. Mode Selection

The mode selector automatically chooses the best mode:

1. Native Async (if Claude Code 2.0.64+) → 5-10 parallel agents (recommended)
2. File-based (always available) → 2 sequential agents (fallback)

3. Coordinator Starts

Native Async Mode:

• Main agent becomes coordinator
• Spawns background agents via Task(run_in_background: true)
• Polls progress via TaskOutput(block: false)
• Manages phase transitions

4. Agents Work in Parallel

Each background agent:

• Receives its subtask and phase
• Works independently
• Shares insights via .claude/popkit/insights.json
• Returns results when complete

5. Insight Sharing

Agents share discoveries via JSON file:

{
  "insights": [
    {
      "agent": "code-explorer",
      "content": "Found existing User model at src/models",
      "tags": ["model", "database"],
      "phase": "explore"
    }
  ]
}

6. Sync Barriers

Between phases:

• Coordinator waits for all TaskOutput to complete
• Aggregates results from phase
• Spawns next batch of agents
• Process continues until all phases complete

7. Completion

When objective met:

• Coordinator aggregates all results
• Patterns saved for future
• State cleared from .claude/popkit/power-state.json

Communication Channels

Uses file-based communication:

File	Purpose
.claude/popkit/insights.json	Shared discoveries
.claude/popkit/power-state.json	Session state

Guardrails

Automatic protections:

• Protected paths (.env, secrets, .git)
• Human-required actions (deploy, delete prod data)
• Drift detection (working outside scope)
• Unconventional approach detection

Human escalation triggers:

• Modifying security configuration
• Pushing to main/production
• Accessing credentials
• Bulk file deletions
• Actions that might be "cheating"

Example Workflow

User: /popkit:power-mode "Build a REST API with user authentication"

[Coordinator starts, creates objective]

Phase 1: EXPLORE (parallel)
├── code-explorer → Analyzes existing codebase
├── researcher → Researches auth best practices
└── architect → Reviews project structure

[Insights shared via Redis]
explorer → "Found existing User model at src/models"
researcher → "JWT recommended for stateless auth"
architect → "Routes pattern: src/routes/{feature}/"

[SYNC BARRIER - wait for all]

Phase 2: DESIGN (coordinator routes insights)
└── code-architect → Designs auth system
    (receives all Phase 1 insights automatically)

[SYNC BARRIER]

Phase 3: IMPLEMENT (parallel with check-ins)
├── rapid-prototyper → Implements endpoints
│   ├── Check-in 1: "Created login route"
│   ├── Check-in 2: "Added JWT generation"
│   └── Check-in 3: "Need: User model location"
│       ← Receives: "src/models/user.ts" from explorer
│
├── test-writer → Writes tests in parallel
│   └── Check-in: "Testing login endpoint"
│
└── docs-maintainer → Updates documentation
    └── Check-in: "Documenting API endpoints"

[SYNC BARRIER]

Phase 4: REVIEW
└── code-reviewer → Reviews all changes
    (receives aggregated results from all agents)

[Complete - results aggregated, patterns saved]

Configuration

Edit power-mode/config.json:

{
  "intervals": {
    "checkin_every_n_tools": 5,
    "heartbeat_seconds": 15
  },
  "limits": {
    "max_parallel_agents": 6,
    "max_runtime_minutes": 30
  },
  "guardrails": {
    "protected_paths": [".env*", "**/secrets/**"],
    "drift_detection": { "enabled": true }
  }
}

Deactivation

# Use command (recommended)
/popkit:power stop

# Or clear state file
rm .claude/popkit/power-state.json

Troubleshooting

Native Async Mode

Claude Code version too old:

• Update Claude Code to 2.0.64+
• Check version: claude --version
• Power Mode will fall back to file mode

Background agents not spawning:

• Verify you have Premium/Pro tier
• Check .claude/popkit/power-state.json for errors
• Free tier is limited to file-based mode

Insights not sharing:

• Check .claude/popkit/insights.json exists
• Verify file permissions

General

Drift alerts:

• Agent working outside scope boundaries
• Update boundaries or reassign task

Integration

Works with:

• All popkit agents (they gain check-in capability)
• Existing skills (run within power mode context)
• Output styles (check-ins use power-mode-checkin format)

Coordinator agent:

• power-coordinator - Can be invoked as coordinator

Related skills:

• pop-subagent-dev - Single-session alternative
• pop-executing-plans - Parallel session alternative