parallelize
Identify parallelization opportunities for implementation work. Use when planning task execution, designing workflows, or maximizing throughput.
When & Why to Use This Skill
The Parallelize skill is designed to optimize agentic workflows by identifying independent tasks that can be executed concurrently. By analyzing dependency graphs and mapping work units, it enables the system to spawn subagents or trigger parallel tool calls, significantly reducing total execution time and maximizing throughput for complex implementation projects.
Use Cases
- Workflow Optimization: Analyzing a multi-step implementation plan to identify tasks with no mutual dependencies that can be executed simultaneously to reduce latency.
- Subagent Delegation: Designing a strategy to split a large-scale project into independent modules, allowing multiple subagents to work on different components at the same time.
- Concurrent Data Retrieval: Batching multiple file reads or API calls into a single parallel operation to speed up the information gathering phase of a task.
- Throughput Maximization: Identifying and restructuring sequential bottlenecks in a pipeline into a parallelized architecture to improve the overall processing speed of the agent.
| name | parallelize |
|---|---|
| description | Use when planning parallel execution of independent tasks via subagents or concurrent tool calls; for workflow planning or throughput optimization. |
| allowed-tools | ["read_file", "update_plan"] |
| short-description | Analyze parallelization opportunities |
Parallelization Analysis
Read this reference:
references/parallelization.md- Concurrent execution patterns and strategies
Instructions
Important: This is about parallelizing implementation work (spawning subagents, concurrent tool calls), NOT writing parallel code.
Analysis Steps
Identify independent work units
- Which tasks have no dependencies?
- What can be done simultaneously?
- Where are the blocking dependencies?
Map dependency graph
- Draw (mentally or explicitly) dependencies
- Identify critical path (longest sequential chain)
- Find parallelizable branches
Calculate potential speedup
- How much work can run concurrently?
- What's the theoretical maximum speedup?
- What are the practical constraints?
Design execution strategy
- Spawn multiple subagents for independent work
- Use parallel tool calls where possible
- Batch independent operations
- Minimize sequential bottlenecks
Parallelization Patterns
Pattern 1: Independent Reads
Read file A, Read file B, Read file C in parallel
→ Process results sequentially if needed
Pattern 2: Independent Analysis
Analyze module A, Analyze module B in parallel
→ Combine results
Pattern 3: Pipeline Parallelism
Stage 1 (Task A) → Stage 2 (Tasks B, C in parallel) → Stage 3 (Task D)
Dispatching Parallel Work (Codex-compatible)
When tasks are independent, batch tool calls in a single message or use multi_tool_use.parallel for concurrent reads/analysis. Keep scopes narrow and independent to avoid conflicts.
Per AGENT.md
"Always use subagents wherever possible to speed up work" - apply this principle throughout.
Arguments
Work to analyze for parallelization: ${ARGUMENTS}