Parallel Agent Orchestration

# Parallel Agent Orchestration

Part of [Agent Skills™](https://github.com/itallstartedwithaidea/agent-skills) by [googleadsagent.ai™](https://googleadsagent.ai)

## Description

Parallel Agent Orchestration is the discipline of dispatching, coordinating, and aggregating results from multiple concurrent subagents to dramatically accelerate complex tasks. Sequential single-agent execution is the default mode for most AI workflows, but it leaves enormous performance on the table. When a task can be decomposed into independent subtasks — analyzing multiple campaigns, reviewing multiple files, searching multiple data sources — parallel dispatch can reduce wall-clock time by 3-10x while maintaining result quality.

This skill encodes the subagent orchestration patterns developed for the Superpowers extension ecosystem and deployed in production at [googleadsagent.ai™](https://googleadsagent.ai), where Buddy™ routinely dispatches parallel subagents to analyze different aspects of a Google Ads account simultaneously. One subagent analyzes bidding strategy, another evaluates keyword performance, a third assesses creative quality — all running concurrently. The orchestrator then aggregates their findings into a unified recommendation set, resolving any conflicts between the independent analyses.

The key challenges in parallel orchestration are task partitioning (decomposing the work into truly independent units), result aggregation (combining outputs that may conflict or overlap), resource management (respecting rate limits and cost budgets across parallel agents), and progress monitoring (tracking multiple concurrent streams without losing visibility).

## Use When

- A task naturally decomposes into 3+ independent subtasks
- Wall-clock time is a critical constraint (user waiting, SLA requirements)
- Multiple data sources or documents need analysis simultaneously
- Code review spans many files that can be reviewed independently
- Batch operations (migrations, refactoring) across multiple files or services
- You need diverse perspectives on the same problem (ensemble reasoning)

## How It Works

```mermaid
graph TD
    A[Complex Task] --> B[Task Decomposer]
    B --> C[Subtask 1]
    B --> D[Subtask 2]
    B --> E[Subtask 3]
    B --> F[Subtask N]
    
    C --> G[Subagent 1]
    D --> H[Subagent 2]
    E --> I[Subagent 3]
    F --> J[Subagent N]
    
    G --> K[Result Aggregator]
    H --> K
    I --> K
    J --> K
    
    K --> L{Conflicts?}
    L -->|Yes| M[Conflict Resolver]
    L -->|No| N[Unified Result]
    M --> N
    
    O[Resource Monitor] --> G
    O --> H
    O --> I
    O --> J
```

The orchestrator receives a complex task and decomposes it into independent subtasks using a task decomposer (either rule-based for well-known patterns or model-assisted for novel tasks). Each subtask is dispatched to a subagent that executes independently, with a resource monitor enforcing shared rate limits and budget constraints. As subagents complete, their results flow to the aggregator, which merges outputs and detects conflicts. Conflicting results (e.g., subagent 1 recommends increasing bids while subagent 2 recommends decreasing them) are resolved by a conflict resolver that applies domain rules or escalates to the orchestrating agent for judgment.

## Implementation

**Task Decomposer:**

```typescript
interface SubTask {
  id: string;
  description: string;
  context: Record<string, unknown>;
  dependencies: string[];
  priority: number;
}

class TaskDecomposer {
  decompose(task: string, context: Record<string, unknown>): SubTask[] {
    const patterns: Record<string, (ctx: Record<string, unknown>) => SubTask[]> = {
      account_audit: (ctx) => {
        const campaigns = ctx.campaigns as str