Summary
Cole Medin walks through Claude Code’s experimental agent teams feature, showing four Claude Code instances collaborating in real time on a codebase review — with the lead agent autonomously deciding team composition and spinning up Tmux terminals for each member. Unlike sub-agents, which execute in parallel but remain isolated and report only summaries back to an orchestrator, agent teams share a live task list and can message each other directly, enabling genuine coordination across concurrent workers.
Setup requires enabling an environment variable or adding a configuration flag to Claude Code’s settings.json, plus installing Tmux or iTerm2 for the split-pane terminal view. Medin covers the practical ceiling of the feature by referencing Anthropic’s own demonstration: 16 agents collaborating to build a complete C compiler from scratch, at approximately $20,000 in API costs — an extreme example that illustrates both the power and the token-heavy nature of the architecture.
The video includes a custom Claude Code command (a skill template) Medin built to improve how the lead agent coordinates team behavior, compensating for cases where Claude’s default approach underutilizes the feature. A clear breakdown of when to choose agent teams over sub-agents — sub-agents remain preferable for context-isolated tasks where cross-agent messaging isn’t needed — makes this a practical decision guide for anyone already using Claude Code for parallel workloads.
📺 Source: Cole Medin · Published February 09, 2026
🏷️ Format: Tutorial Demo
