How does Exploring in Parallel differ from sequential searches?

It launches multiple subagents at once, reducing wall-clock time and surfacing diverse results to synthesize.

When should you avoid parallelization?

For dependent searches, a single specific file, simple grep tasks, or when the exact target is already known.

What triggers parallelization?

Research, exploration, debugging, why does this happen, how does X work, or multi-file investigations.

exploring-in-parallel

npx machina-cli add skill oryanmoshe/agent-skills/exploring-in-parallel --openclaw

Files (1)

SKILL.md

2.9 KB

Exploring in Parallel

Overview

Parallelize exploration aggressively. When researching or exploring, launch multiple subagents simultaneously instead of sequential searches. This reduces wall-clock time proportionally to the number of parallel agents.

How to Parallelize

Use the Task tool with subagent_type=Explore. The key is to send multiple Task calls in a single message — this launches them concurrently.

Each subagent gets:

subagent_type: "Explore" for codebase searches, "general-purpose" for web research
description: Short label (3-5 words)
prompt: Detailed, self-contained search instructions

Example — investigating how auth works:

Send in ONE message:

Task 1: { subagent_type: "Explore", description: "Find auth middleware", prompt: "Search for authentication middleware, auth guards, session validation..." }
Task 2: { subagent_type: "Explore", description: "Find login endpoints", prompt: "Search for login routes, /auth endpoints, credential handling..." }
Task 3: { subagent_type: "Explore", description: "Find auth config", prompt: "Search for JWT config, OAuth settings, auth environment variables..." }

All three run simultaneously. When results return, synthesize findings into a coherent answer.

When to Parallelize

Situation	Action
Searching for multiple patterns	One subagent per pattern
Exploring different directories	One subagent per area
Investigating related questions	One subagent per question
Checking multiple files (>3)	Parallel reads
Web research on multiple topics	One subagent per topic

When NOT to Parallelize

Dependent searches: Second search needs results from first
Single specific file: Just use Read directly
Simple grep: One pattern, one location — use Grep directly
Known target: Know exact file/function — use Glob/Read directly

Red Flags — STOP and Parallelize

Thought	Action
"Let me search for X first, then Y"	Launch both simultaneously
"I'll check this file, then that one"	Parallel reads if >3 files
"First explore area A, then B"	Parallel subagents per area
"Let me see what this returns before..."	Usually can parallelize anyway

After Results Return

Don't just dump raw results. Synthesize:

Identify overlapping findings across agents
Resolve contradictions
Present a unified answer with file references
Note any gaps that need follow-up

Source

git clone https://github.com/oryanmoshe/agent-skills/blob/main/skills/exploring-in-parallel/SKILL.mdView on GitHub

Overview

Exploring in Parallel accelerates codebase research by firing multiple subagents at once to search across files, patterns, and sources. This approach reduces wall-clock time and surfaces diverse context for deeper understanding. Use it when you need fast, multi-source exploration.

How This Skill Works

Use the Task tool with subagent_type=Explore and send several Task calls in a single message. Each subagent gets a short description, a type, and a detailed prompt. Results from all subagents are then synthesized into a coherent, unified answer.

When to Use It

Searching for multiple patterns across the codebase
Exploring different directories or modules
Investigating related questions or edge cases
Checking multiple files (more than 3) in parallel
Web research on multiple topics or sources

Quick Start

Step 1: Define goals and the areas or patterns to search
Step 2: Send a single message with multiple Task calls, each including subagent_type, description, and prompt
Step 3: Collect results, synthesize findings, and present a unified answer

Best Practices

Define the patterns or questions upfront and assign one subagent per pattern or area
Parallelize when you genuinely need multi-file or multi-source results (more than 3 files)
Keep subagent prompts concise with a clear 3–5 word description and focused instructions
Always synthesize results and resolve contradictions before proceeding
Tag results with file references and note gaps for follow-up

Example Use Cases

Audit a codebase to trace authentication flow across modules
Investigate a bug by searching across config, middleware, and routes
Research how a feature is implemented in frontend and backend across files
Compare implementations of a pattern across multiple repositories
Gather context from docs, tests, and code for a security finding

Frequently Asked Questions

Add this skill to your agents