Claude Code is cooking

Quick one today but a good one if you're starting to experiment with AI in your dev workflow, Claude Code got a major upgrade “sub-agents” last week. And… they seem like a big deal.

So what are Claude sub-agents?

Think of them as little expert teammates inside Claude Code. Each with their own specialty, tools and memory. Instead of asking a general purpose AI to do everything, sub-agents let you spin up task specific agents

• A git manager for your repo workflows

• A debugging agent for tracing issues

• A test writer or a doc generator

Each sub-agent has its own 200k token context window, so they don't pollute the main convo or forget what they're doing halfway through. They also don't talk directly to you. The main Claude Code agent orchestrates the flow.

How does it work?

1. You send a prompt to the primary Claude agent

2. That agent decides which sub-agent (if any) to activate

3. The sub-agent does its thing in isolation and returns the result

4. The primary agent packages it all back up for you

So when you create a sub-agent, you're essentially defining its system prompt, tools and when it should be used. That last part “when to use it” is key. It helps the main agent know when to call it.

Why It's Useful

• Sub-agents work in their own space, so you don't hit token limits as quickly and hallucinations drop significantly.

• Each sub-agent is built for a narrow task, which makes it more reliable.

• You can store them in your repo and reuse across projects or teams.

• You can limit which tools (like MCPs or other plugins) each agent can use.

Honestly, this feels like the closest thing we've had to a real "AI project manager" in the terminal. And one that can delegate intelligently.

Some pitfalls

• No memory: Sub-agents only know what the main agent tells them. Don't assume shared context.

• Debugging is tricky: You don't see full logs, just tool call results.

• Don't go overboard: If you add too many sub-agents with vague descriptions, the main agent may get confused.

Some Cool Use Cases

• A design system enforcer that reviews UI code

• A delegator that breaks up a feature and assigns it to other sub-agents

• A meta agent that builds more sub-agents (this one's wild)

Industry Pulse

This represents a massive shift toward multi-agent orchestration becoming mainstream. While tools like CrewAI and AutoGen pioneered the concept, Claude Code makes it accessible to every developer without complex setup or API management.

The timing is perfect though. As context windows hit practical limits, intelligent task delegation becomes essential for maintaining AI effectiveness at scale.

Bottom line: If you're building with Claude Code (or just dipping your toes in agentic dev), sub-agents are worth trying. They let you scale yourself a little, without turning everything into chaos.

Let me know if you build any cool ones. I'm collecting examples.

👀 Whats shipping this week?

• NotebookLM Video Overviews - Google's tool transforms complex source material into AI generated videos, potentially useful for developers creating technical documentation or educational content from codebases

• Terragon background agents for Claude Code - Background agent system that enables firing off multiple Claude Code agents in parallel, particularly useful for mobile development and complex multi-task automation.

• ElevenLabs CLI for Conversational AI Agents - CLI that allows developers to manage conversational AI agents as code with version control, automation, multi-environment configurations and secure authentication for voice and chat workflows.

📖 Worth the scroll

• How AI Vibe Coding Is Destroying Junior Developer's Careers - building applications entirely with AI without learning to code is creating pseudo developers who can generate code but can't debug, understand or maintain it making them vulnerable in an already brutal job market where fundamental programming skills remain essential.

🤓 Case Studies

• Intercom's three lessons for creating a sustainable AI advantage - Intercom transformed from experimenting with GPT-3.5 within hours of release to launching Fin AI Agent just 4 months later, now resolving millions of customer queries monthly.

• Exa built a production multi-agent research system using LangGraph that processes hundreds of daily queries in 15 seconds to 3 minutes, featuring dynamic task generation, structured JSON outputs for API consumption, and LangSmith observability for token tracking and cost optimization.

💬 Quick question: What's the most time consuming part of your development workflow? Reply and I’ll build a tutorial to automate it.

📱 Stay connected: Follow me on LinkedIn and Twitter/X for daily AI tool breakdowns and quick wins

Thanks for reading
- Sanket