Claude’s Custom Agent Creator (Wizard) Update
Anthropic accidentally leaked 512,000 lines of Claude Code source code on March 31st, 2026, revealing a custom agent creator wizard that allows users to build specialized AI sub-agents. The leak exposed unreleased features including the Chyros background daemon, a three-layer memory system, and the full custom sub-agent orchestration platform.
Summary
On March 31st, 2026, researcher Chiao Fanhaw discovered that Anthropic had accidentally leaked their Claude Code source code through a routine update to version 2.188 on the NPM registry. The leak contained 512,000 lines of TypeScript across nearly 2,000 files, caused by human error when a debug file was included in the public release. The post about the leak received over 21 million views and was forked over 10,000 times on GitHub before Anthropic confirmed the incident and implemented measures to prevent future occurrences.
The leaked code revealed Claude Code's agentic harness - the software layer that manages tools, memory, and tasks - rather than the actual model weights or training data. Key discoveries included feature flags for unreleased capabilities, a background daemon called Chyros that maintains work during idle periods, a three-layer memory system, internal model code names, and a hidden Tamagotchi-style virtual pet.
The most significant discovery was the custom sub-agent system with an interactive wizard accessed through the 'arch agents' command. This system allows users to create specialized AI workers with configurable parameters including name/description (for task routing), model selection (Sonnet for efficiency, Opus for complex reasoning), tool access restrictions, memory persistence, and scope settings. Sub-agents solve context window limitations by spawning separate processes with clean contexts for specific tasks, moving beyond Claude Code's original single-session approach.
The leaked code also revealed future developments including Chyros for always-on background processing and agent teams (already shipped experimentally with Opus 4.6) that enable multiple Claude sessions to work as coordinated teams. The speaker positions this evolution as transforming Claude Code from a terminal assistant into a multi-agent orchestration platform, emphasizing that users who master custom agent creation will have significant productivity advantages.
Key Insights
- Anthropic accidentally leaked 512,000 lines of Claude Code source code on March 31st, 2026 due to human error when a debug file was included in a routine NPM registry update, receiving over 21 million views
- The leaked source code revealed an unreleased background daemon called Chyros that keeps Claude working while users are idle, consolidating memory and preparing for next sessions
- Claude Code's sub-agent system solves context window limitations by letting Claude spawn separate child processes with their own clean context windows, tools, and instructions for specific tasks
- The custom agent wizard allows users to configure agents with specific models (Sonnet for efficiency, Opus for complex reasoning), tool restrictions, persistent memory, and scope settings that become immediately available after saving
- Agent teams already shipped as an experimental feature with Opus 4.6 in February 2026, enabling multiple Claude sessions to work as coordinated teams with direct messaging through shared task lists
Topics
Transcript
[0:00] 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 [0:31] smarter and build better processes. And this story is one of the biggest AI news moments of 2026…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Julian Goldie SEO
NEW Nvidia Autonomous AI is WILD!! 🤯
Nvidia announced Nemo Clo, a new autonomous AI agent system that operates independently without continuous prompting. Powered by Nemotron 3 Ultra (a 550 billion parameter model), the system is five times faster and cheaper than previous versions, with OpenShell providing secure sandboxed execution.
Laguna XS 2.1: New FREE + Opensource Local AI!
Julian reviews Laguna XS 2.1, a new free open-source local AI coding model from Poolside that performs comparably to Qwen 3.6 and outperforms Claude Haiku on benchmarks. He demonstrates its practical capabilities by building landing pages and functional apps, highlighting its speed, offline functionality, and multiple deployment options through local setup, Claude Code, or OpenRouter's free API.
How to Run Hermes FREE Forever!
The video demonstrates how to run the Hermes AI agent for free using Gemma 4, a local open-source model from Google, with significant speed improvements through MLX optimization. The setup works on Apple Silicon Macs or via free APIs on Open Router, enabling autonomous agents to work offline and privately without subscription costs.
This NEW Chinese AI is INSANE! (FREE + Open Source!)
Long Cap 2.0 is a new open-source Chinese AI model from a food delivery app company that offers 1 million tokens of free context memory, beats GPT-4.5 on SWE bench pro benchmarks, and uses efficient parameter activation to reduce computational overhead while maintaining high performance.
Claude Code is now FREE: Here’s how…
Google's new Gemma 4 model running on Ollama is 90% faster on Apple Silicon, enabling free Claude Code usage locally without token costs. The setup requires three simple steps: downloading Ollama, Gemma 4, and installing into Claude Code, with alternatives available via OpenRouter API for non-Mac users.