Descriptions:
Web Dev Cody walks through Claude Code’s built-in subagent system, demonstrating how to offload research and implementation tasks to parallel agents that each maintain their own separate context windows. The core problem he addresses is that running work through the main Claude Code thread fills up context quickly — a simple security audit scan pushed the window to 30% — which degrades output quality over time. Subagents sidestep this by doing focused work in isolation and reporting results back to the main thread without polluting it.
The tutorial covers creating custom agents inside the /agents menu, including choosing models per agent (Haiku for fast file-scanning tasks, Opus for code implementation work), setting tool permissions, and dispatching multiple agents concurrently. Cody shows five FileFinder agents running in parallel to scan a codebase for authentication-related files, noting that parallelism produces more thorough coverage than a single sequential search because each agent has its own fresh context.
Practical takeaways include designing a two-agent workflow — a lightweight Haiku-powered researcher paired with an Opus-powered “clean code architect” — and using the subagent pattern as a general strategy to keep the main thread lean during long agentic sessions. The video is aimed at developers already using Claude Code who want to scale their workflows without hitting the context ceiling.
📺 Source: Web Dev Cody · Published December 24, 2025
🏷️ Format: Tutorial Demo
