Claude’s Custom Agent Creator (Wizard) Update

Key Takeaways

Claude's custom agent creator wizard update, what the leaked source code reveals. Something huge just happened. Anthropic accidentally leaked their own source code, half a million lines, and buried inside it a feature most people don't even know exists yet. A custom agent creator, a wizard that lets you build your own AI agents from scratch inside Claude code. This changes how you work with AI, and today we're breaking down exactly what it is and how to use it. Hey, I'm the digital avatar of Julian Goldie. I help people learn and implement AI tools so they can work smarter and build better processes. And this story is one of the biggest AI news moments of 2026 so far. So, let's get into it. We've got the leak itself, what Claude code is if you're new to this, and then the real gem, the custom agent creator, how it works, and how you can start using it right now. Stay with me because the stuff at the end is the most practical part. On March 31st, 2026, a researcher named Xiao Fan Show spotted something weird, a routine update to Claude code version 2.18.8 had been pushed to the NPM registry. That's the public package database developers use to download software. It was 60 megabytes heavier than it should have been. Inside that extra weight, a source map file, a debug file that lets you reconstruct the original code from a compressed version. These are meant to be internal. They should never ship to users. This one did. Show posted a download link on X. Within hours, the post got over 21 million views. Around 512,000 lines of TypeScript across nearly 2,000 files was mirrored across GitHub and forked over 41,000 times. Anthropic confirmed it. A release packaging issue caused by human error. No customer data, no credentials exposed. They've been rolling out measures to prevent it happening again. This isn't even the first time. Back in February 2025, an earlier version of Claude code leaked the same way. This time it was bigger, more detailed, and came right after a separate leak involving an internal blog post about an unreleased model. The leaked source doesn't contain Claude's model weights or training data. It contains the agentic harness, the software layer that tells Claude how to use tools, manage memory, and run tasks. And inside that harness, developers found some really interesting things, feature flags for unreleased capabilities, a background daemon called Kairos that keeps Claude working while you're idle. A three-layer memory system, internal model code names, even a hidden Tamagotchi-style virtual pet set to drop on April 1st. Draw your own conclusions there. But the most important discovery for anyone using AI to improve their work is the custom sub-agent system, and specifically the wizard that lets you build and configure your own agents. Let me explain what sub-agents are because this is where it gets useful. Claude code is Anthropic's command-line tool built on top of Claude. You give it complex instructions, refactor this code base, fix this bug, and run the tests, and it works through the task autonomously. Normally, there's one context window, one thread. Claude reads everything, thinks, responds, but context windows have limits. Long sessions get messy. Sub-agents solve this by letting Claude spawn separate child processes for specific tasks. Each gets its own clean context window, its own tools, its own instructions. Claude delegates instead of trying to do everything in one thread. Claude code already ships with built-in sub-agents. An explore agent for searching code bases, a plan agent for research before making changes, a general-purpose agent for complex multi-step work. These run automatically in the background, but you can build your own. That's what got people excited. Inside Claude code, there's a command called alt agents. Run it and you get an interactive wizard to view, create, edit, and manage your sub-agents. When you create a custom agent, here's what you configure. The name and description. This is critical. It's how Claude decides when to delegate to your agent. Write a clear, specific description, and Claude automatically routes the right tasks to it. The model. Set it to Sonnet for fast, efficient work. Use Opus when you need heavier reasoning. Different models for different jobs. Tool access. Want a read-only research agent? Limit it to read, grep, and blob. No write access. That constraint is intentional. Your research agent can never accidentally modify files. Memory. Give the agent a persistent memory directory, and it carries insights across conversations, code base patterns, recurring issues, things it's learned. Or set it to none, and it starts fresh each time. Scope. User scope makes the agent available across all your projects. Project scope keeps it to one specific project. Hit save, and the agent is live immediately. Just type use the agent name agent to, and Claude delegates. Agents are stored as markdown files with a YAML config block in a .claude/agents directory, so you can also create and edit them manually if you prefer. When I first started with Claude code, it was overwhelming. So many features, so many ways to configure things. And that's part of why I built a community called AI Profit Boardroom, over 2,000 members all focused on learning AI together and sharing what actually works, real use cases, practical implementations, workflows that save time versus ones that waste it. If you're serious about using AI to improve your work and skills, check it out. Link in the description. Let's talk about where this is headed because the leaked code gives us a clear picture. Early Claude code was a single-session tool, smart but one thread. The sub-agents added parallelism and context isolation. Now, custom agents let you define your own specialized workers tailored to your exact workflow. What's coming next is even bigger. Kairos is an always-on background mode that keeps working while you're idle, consolidating memory, merging observations, preparing for your next session. And agent teams already shipped as an experimental feature with the Opus 4.6 release in February 2026. Multiple Claude sessions working as a coordinated team. One lead, multiple teammates messaging each other directly through a shared task list. True parallel collaboration between AI workers, not just isolated sub-agents reporting back. Claude code has gone from smart terminal assistant to multi-agent orchestration platform, and it's still moving fast. Here are some practical tips for getting the most out of custom agents. Write specific descriptions. Analyze this Python files for code quality and suggest improvements will trigger correctly. Code reviewer probably won't. Use tool restrictions intentionally. A research agent with read-only access is safer and more predictable. Match permissions to what the agent actually needs. Match model to task. Sonnet for fast, frequent tasks. Opus for complex reasoning. Mixing models across agents is a real efficiency lever. Enable persistent memory for agents you'll use repeatedly. A code quality agent that remembers your code base gets more useful over time. Share agents with your team. Claude code supports distributing sub-agents through plugins. Build it once, share across your project without everyone starting from scratch. And check out the awesome Claude code sub-agents repository on GitHub. It's a curated list of ready-made sub-agent definitions you can drop straight into your project. Here's the bottom line. Custom agents let you build a specialized AI workforce tailored to your exact needs. A security reviewer that knows your code base, a documentation agent that follows your style guide, a testing agent scoped only to your test files. You're not limited to what Anthropic ships by default. You're configuring your own system. The people who learn to build and direct these agents well are going to be significantly more productive than those who don't. That's just where things are going. To recap, Anthropic accidentally leaked 512,000 lines of Claude code source code on March 31st, 2026. Human error, no customer data exposed. The leak revealed unreleased features including Kairos, a three-layer memory system, and the full picture of the custom sub-agent system. Claude code's agent wizard lets you create custom agents with specific models, tools, memory, and scope available immediately after saving. If you're looking to dive deeper into AI tools and actually implement them in your work, I recommend AI Profit Boardroom, over 2,000 people learning how to use AI effectively. Everyone shares real experiences. What's working, what's not, which tools are worth your time, no hype, just practical guidance from people doing the work. Link in the description. And if you want the full process, SOPs, and 100+ AI use cases like this one, join the AI Success Lab. Links in the comments and description. You'll get all the video notes from there, plus access to our community of 58,000 members who are crushing it with AI. If this was useful, hit like, subscribe, and I'll see you in the next one.