ls-max-parallel

name	ls-max-parallel
description	Use when decomposing a task into parallel subtasks with structured subagent spawning. Triggers on 'max parallel', 'parallelize this', 'spawn subagents', 'decompose and parallelize', 'break this into parallel tasks', 'recursive decomposition'. NOT for simple agent swarms (use /max-swarm).
allowed-tools	Task, TaskCreate, TaskUpdate, TaskList, Read, Glob, Grep

Maximum Parallelism Mode Activated

You are now in MAX PARALLEL mode. Your behavior changes:

Flag Parsing

Parse the user's input for these flags (apply defaults for any not specified):

Flag	Default	Effect
--agents N	10	Max agents per wave
--depth N	1	Recursive decomposition depth (2 = sub-agents can decompose further)
--ensemble	off	Multiple models tackle same subtask, then merge results
--strategy S	recursive	Strategy: recursive, ensemble, sweep

Strategies

recursive (default)

Decompose -> Execute in waves -> Synthesize. Standard parallel decomposition with high agent count.

ensemble

Same subtask given to 2-3 different models independently, then merge/compare results. Best for high-stakes decisions where confidence matters more than speed.

Wave 1 (ensemble - same question, different models):
  - Task("Analyze auth architecture", model="haiku")
  - Task("Analyze auth architecture", model="sonnet")
  - Task("Analyze auth architecture", model="opus")

Wave 2 (synthesis):
  - Task("Merge and compare the 3 analyses", model="opus")

sweep

Broad exploration -- maximize breadth in Wave 1 (up to 10 agents), each exploring a different angle. Best for codebase understanding and open-ended research.

Wave 1 (sweep - 10 agents, each a different angle):
  - Task("Explore auth patterns", model="haiku")
  - Task("Explore database layer", model="haiku")
  - Task("Explore API routes", model="haiku")
  - Task("Explore error handling", model="haiku")
  - Task("Explore test patterns", model="haiku")
  - Task("Explore config/env setup", model="haiku")
  - Task("Explore logging/observability", model="haiku")
  - Task("Explore state management", model="haiku")
  - Task("Explore build/deploy pipeline", model="haiku")
  - Task("Explore dependency graph", model="haiku")

Wave 2 (synthesis):
  - Task("Synthesize all 10 exploration results", model="opus")

Immediate Actions

1. Parse flags from user input (apply defaults for missing flags)
2. Select strategy based on --strategy flag
3. Analyze the current task or last user request
4. Decompose into ALL subtasks (target --agents count per wave)
5. Map dependencies between tasks
6. Spawn agents in parallel for all independent work

Execution Pattern

Step 1: Full Task Decomposition

List EVERY subtask needed to complete the request. Target up to 10 subtasks per wave (or --agents N). Don't stop at 2-3 -- find ALL components.

Step 2: Dependency Mapping

For each task pair, ask: "Does Task A need Task B's output?"

• YES -> Task B blocks Task A
• NO -> They can run in parallel

Step 3: Wave Grouping

• Wave 1: All tasks with no blockers (run in parallel NOW -- up to 10 agents)
• Wave 2: Tasks blocked only by Wave 1 (run after Wave 1 completes)
• Wave 3+: Continue pattern
• Wave N+1 (Synthesis): MANDATORY final wave -- synthesize and validate all results

Step 4: Parallel Execution

Spawn ALL Wave 1 agents in a single message with multiple Task tool calls.

Task("Subtask A", subagent_type="Explore", model="haiku")
Task("Subtask B", subagent_type="Explore", model="haiku")
Task("Subtask C", subagent_type="Explore", model="haiku")
Task("Subtask D", subagent_type="general-purpose", model="sonnet")
Task("Subtask E", subagent_type="general-purpose", model="sonnet")
Task("Subtask F", subagent_type="Explore", model="haiku")
Task("Subtask G", subagent_type="Explore", model="haiku")
Task("Subtask H", subagent_type="general-purpose", model="sonnet")
Task("Subtask I", subagent_type="Explore", model="haiku")
Task("Subtask J", subagent_type="Explore", model="haiku")

Step 5: Wave Progression

After Wave 1 completes, immediately spawn ALL Wave 2 agents in parallel. Continue until all execution waves complete.

Step 6: Synthesis Wave (MANDATORY)

After all execution waves, spawn a synthesis agent to merge and validate results:

Task("Synthesize and validate all results from Waves 1-N", subagent_type="general-purpose", model="opus")

The synthesis agent must:

• Merge findings from all agents into a coherent summary
• Flag any contradictions or conflicts between agent results
• Validate completeness -- did we miss anything?
• Produce the final deliverable

Depth Control (--depth)

If --depth >= 2, execution agents may themselves decompose subtasks into further parallel waves. Each sub-agent follows the same pattern: decompose -> map -> execute in waves -> synthesize. Max recursion = --depth value.

Model Selection

Task Complexity	Model
File finding, globbing, simple grep	haiku
Code reading, analysis	sonnet
Complex reasoning, architecture	opus
Ensemble verification (same task, different model)	mix haiku + sonnet + opus
Synthesis / final merge	opus

The Iron Law

If tasks are independent, they MUST run in parallel. Sequential execution of independent tasks is a bug.

Quick Reference

Examples

# Default: 10 agents, recursive strategy, depth 1
/max-parallel Add user authentication

# 6 agents, sweep strategy for broad exploration
/max-parallel --agents 6 --strategy sweep Explore this codebase

# Ensemble mode for a critical decision
/max-parallel --ensemble Should we use Redis or PostgreSQL?

# Deep recursive decomposition (sub-agents can decompose further)
/max-parallel --depth 2 --agents 8 Refactor the entire auth system

# Conservative: fewer agents for a smaller task
/max-parallel --agents 4 Fix all lint errors

Execution Flow

1. Parse flags (--agents, --depth, --strategy, --ensemble)
2. Select strategy
3. Decompose task into subtasks (target: --agents per wave)
4. Map dependencies between subtasks
5. Group into parallel waves
6. Execute waves (Wave 1 -> Wave 2 -> ... -> Wave N)
7. Synthesis wave (MANDATORY) -- merge and validate all results

NOW: Apply to Current Context

What task should I parallelize? If you have a pending request, I will:

1. Parse any flags from your input
2. Select the appropriate strategy
3. Show the full task decomposition (targeting up to 10 agents per wave)
4. Show the dependency graph
5. Execute Wave 1 immediately with parallel agents
6. Run synthesis wave after all execution waves complete

Awaiting your task or confirm to parallelize the last discussed work.