Problem
AI agents must keep everything in conscious attention - strategy, tactics, implementation, all file context. This cognitive overload degrades performance.
Humans chunk practiced tasks into automatic execution, freeing attention for higher-level thinking. We can't retrain production models, but we can simulate this.
Pattern
Use a main orchestrator agent with focused subagents:
- Main agent stays strategic: Plans, designs, breaks down work, integrates results
- Subagents handle execution: Read files, implement, test
- Delegate in parallel when possible: Multiple subagents on independent chunks
Key insight: Delegate execution to subagents like humans delegate practiced skills to automatic processes. Orchestrator operates strategically while subagents handle details.
How to apply:
- Use planning mode for mid-size tasks
- Request plans for parallel subagent execution
- Include comprehensive test plans
- Configure agent: "Always delegate coding tasks to subagents"
When this works:
- Mid-size tasks in clean, modular codebases
- First prototypes
- Well-defined interfaces between components
When this fails:
- Non-modular code
- Tightly coupled systems
- Subagents make incompatible tradeoffs
Example
Building this website: Task: "Create Next.js app with unit and E2E tests, simple professional design"
Main agent planned architecture and test strategy. Subagents implemented in parallel: markdown processing, page components, unit tests, E2E tests. Result: fully functional site in one iteration.
Multi-layer system: Main agent designed layer interfaces. Subagents worked on separate layers following contracts. Worked because codebase was modular. In poorly structured code, subagents create integration problems.
