Anthropic's Entire Claude Code Roadmap Just Leaked

On March 31, 2026, a security researcher posted a link on X showing that the Claude Code npm package had shipped with a source map file that exposed the full source. Anthropic pulled the old versions later that day, but the contents were already out.

The package contained 512,000 lines of TypeScript across 19,000 files - tools, permission systems, internal code names, the works. It happened because a single source map file got included in a public package by mistake.

The interesting part isn't the leak itself, it's what was in there: roughly 44 unreleased Claude Code features, many of them already built.

I've been teaching Claude Code for months now. I've trained over 100 people to build real software with it. And after reading through the leaked roadmap, my takeaway is simple: agent orchestration is the next big thing for tech professionals, and Claude Code is about to become something much bigger than a coding tool.

The Five Features That Matter Most

1. Chyros: Always-On Autonomous Claude

Right now, when you use Claude Code, you babysit it. You give it a task, watch it work, course correct, repeat. Chyros changes that fundamentally. It's an always-on autonomous system that can run unattended for hours, check its own tasks, and create new tasks for itself.

The part I found most interesting is something the code calls "nightly dreaming." A sub-agent runs four phases (orient, gather, consolidate, prune) and distills the day's work logs into structured knowledge that gets baked into long-term memory overnight.

That's a meaningful shift from AI assistant to something closer to an AI employee. You give it a goal on Monday and it works toward it for the rest of the week.

2. Coordinator Mode: Multi-Agent Swarm Management

Power users already run multiple Claude Code instances in parallel. But as that scales to 50 or 100 agents doing different work, coordination becomes the bottleneck. Coordinator Mode is Anthropic's answer: a strategy-game-style control board for managing swarms of AI agents working together on the same codebase.

When this ships, a lot of the work that currently looks like "hire 10 junior devs" is going to look more like "spin up 10 Claude agents with a coordinator on top."

3. Wizard: Custom Agent Creator

Currently, Claude Code has a fixed personality and toolset defined by Anthropic. Wizard lets you build your own agents with custom personality, custom memory types, custom tool access, and custom deployment locations (desktop, cloud, or local hardware).

Combined with Coordinator Mode, the picture becomes clear: you're not just using one AI assistant. You're building and managing a team of specialized AI agents.

4. Ultra Plan: 30-Minute Cloud Sandbox for Deep Thinking

Anyone who's used Claude Code on a large project knows the context window problem. Yes, it technically supports 1 million tokens. But performance degrades noticeably after 200K. Ultra Plan offloads planning to a 30-minute cloud sandbox with massive context, producing a much better blueprint before Claude starts building.

This solves one of the most common complaints I hear from students: "Claude loses track of what it's doing on complex projects."

5. Buddy: The Gamified AI Companion

This one is pure fun. Buddy is a virtual pet that lives in your CLI. 18 species (ducks, cats, dragons, capybaras), rarity tiers from common to legendary, stats like debugging, patience, chaos, and wisdom. It's a Tamagotchi for developers.

It looks like a gimmick but I think it's actually a smart on-ramp. Gamification is how casual users get pulled into a power tool, and Buddy is probably going to be the reason a lot of non-technical people try Claude Code for the first time.

New Models Confirmed

The leak also confirmed several upcoming models through Anthropic's internal "undercover mode" system (ironically, the very system designed to prevent these leaks):

• Capibara (internal codename): a new model already being used internally to build Claude Code itself
• Mythos: trained on 5-10 trillion parameters, roughly 3x the current models. Reportedly flagged as a cybersecurity risk due to its capabilities
• Claude Opus 4.7 and Sonnet 4.8: versioned upgrades to the current lineup

What This Means If You're Learning Claude Code

I'm biased because I teach this stuff, but for what it's worth: every feature on this roadmap makes Claude Code more powerful and more complex. Multi-agent orchestration, custom agent creation, autonomous long-running agents, advanced planning. These aren't easier versions of what exists today, they're entirely new capabilities layered on top.

I think the people who understand the fundamentals now will absorb these new features pretty quickly. The people who wait will be looking at a steeper learning curve. It's the same pattern that played out with AWS, Kubernetes, and Salesforce - simple tools became platforms, platforms created training ecosystems, and the early practitioners became the experts everyone else paid to learn from.

The Bigger Picture

Claude Code is clearly not a side project for Anthropic. 44 features built and ready to ship, a dedicated team using their most advanced internal models, and internal systems (like undercover mode) specifically designed to protect its secrets. It's the platform play.

And the demand side is matching the build side. Anthropic just shared that their revenue run-rate has crossed $30 billion, up from $9 billion at the end of 2025, with most of that acceleration coming from enterprise. They're also lining up new compute partnerships to keep up. None of that slows down. If anything, it's the reason a roadmap this ambitious exists in the first place - they have the customers to justify shipping all of it.

Six months from now, Claude Code probably isn't going to look like a terminal app that writes code. It's going to look more like an operating system for AI agents, and the people who got familiar with it early are going to be the ones running it.