Claude Code & Custom Agents

Overview

Sal presented the evolution of agentic coding tools, focusing on Claude Code’s approach to context engineering and custom sub-agents. The talk covered how Claude Code addresses LLM limitations through instruction files, custom agents, and parallel execution patterns.

The Evolution: From Prompt to Context Engineering

The Dark Ages of Prompt Engineering (2022-2023)

• Everyone focused on crafting the “perfect prompt”
• Patterns weren’t generalizable across models
• Inefficient compared to modern tooling
• Each model had its own quirks and intricacies

The New Era: Context Engineering

• MCP (Model Context Protocol) and tooling provide scaffolding around foundational models
• Claude Code allows model selection (Sonnet, Opus)
• Focus shifted from prompts to managing context effectively

Solving the Ephemeral Memory Problem

The Challenge

• LLMs are stateless (like REST APIs)
• Context windows get compacted or cleared
• Loss of relevant information over time

The Solution: Instruction Files

• Claude.md files persist state between requests
• LLM-friendly format for maintaining context
• Survives context window compaction/clearing
• Greater control over context compared to cursor/copilot

Custom Sub-Agents in Claude Code

Evolution of Agent Capabilities

1. Task Command: Early ability to run separate tasks

   • Bash commands
   • Web scraping
   • Role-based sub-agents

2. Custom Agents Feature (Released 2-3 months ago)

   • Pre-defined sub-agents with own system prompts
   • Don’t inherit from Claude.md (isolation)
   • Complete tasks independently
   • Defined using /agents command

Key Features of Custom Agents

• Isolated Context Windows: Each agent has its own context
• Parallel Execution: Run multiple agents simultaneously
• Tool Selection: Choose specific tools including MCP servers
• Transportable: Markdown files can move between projects
• Domain-Specific: Specialized for particular tasks

Best Practices for Custom Agent Development

General Guidelines

1. One Job Per Agent: Keep agents focused
2. Start Read-Only: Begin with read tools, add editing as needed
3. Restrict Tools: Minimize context pollution
4. Clear Descriptions: Help orchestrating agent know when to invoke

Example Agent Types

• Code Reviewer: Runs after significant code changes
• QA Agent: Testing and validation
• Debugging Agent: Error analysis and fixes
• Documentation Agent: Generate/update docs

Live Demo: Building an Astro Blog

Sal demonstrated parallel agent execution for creating a personal site with blog functionality:

• Multiple agents running simultaneously
• Each handling specific aspects (components, styling, content)
• Visual demonstration of parallel processing power

Execution Patterns

• Serial Execution: Pass information from agent to agent
• Parallel Execution: Multiple agents work independently
• Phase-Based: Organize agents into execution phases

Context Management Considerations

The Downsides

• Token Cost: Each agent consumes context space
• Context Pollution/Enrichment: Balance between too much and too little
• Compaction Frequency: More agents = more frequent clearing
• Quality Impact: Frequent compaction reduces output quality

Optimization Strategies

• Keep agent prompts concise
• Limit number of large sub-agents
• Monitor context usage with /context command
• Balance between functionality and token efficiency

Q&A Highlights

Q: How well do sub-agents work for visual/multi-modal tasks?

• Mixed results with accessibility testing
• Can use Playwright MCP for headless Chrome screenshots
• Image analysis capabilities vary
• Still a work in progress

Q: How do you determine the atomic unit for an agent?

• Think like database transactions
• Group related small tasks together
• Example: React component creation vs unit testing
• Keep related functionality cohesive

Key Takeaways

1. Context engineering > Prompt engineering for modern LLM applications
2. Custom agents provide isolation and prevent context pollution
3. Parallel execution dramatically speeds up complex tasks
4. Balance is key: Too many agents cause token overhead
5. Claude Code gives fine-grained control over context and execution

Resources

• 📺 Watch Sal’s Claude Code & Custom Agents Talk on YouTube
• Presentation Slides (PDF)
• Claude Code documentation
• MCP (Model Context Protocol) specification
• Community agent repositories (search GitHub)
• /agents command for agent management
• /context command for monitoring token usage

About the Speaker

Sal Castoro is an AI engineer specializing in agentic terminal-based tools and Claude Code implementations. He actively explores the boundaries of context engineering and multi-agent orchestration patterns.