Claude Code Just Became a Full Agent Platform — Here's What Changed (Skills 2.0)

Most people are still using Claude Code like a fancy autocomplete. Meanwhile, Anthropic just shipped what I think is the biggest upgrade since launch — and almost nobody is talking about it.

Skills aren't instructions anymore. They're programs.

Here's what dropped:

🔧 SKILLS SYSTEM (merged with Commands)

- Custom slash commands and skills are now the same thing

- Every skill gets a SKILL.md file with YAML frontmatter — description, tool restrictions, model routing, execution isolation

- Claude auto-loads the right skill based on what you're doing. No manual switching.

- Skills are directories now — templates, examples, validation scripts, all bundled together

4 built-in skills ship out of the box:

- /simplify — spawns 3 parallel review agents to clean your code

- /batch — decomposes large changes into 5-30 independent units, each gets its own agent + PR

- /debug — reads debug logs and troubleshoots your session

- /loop — runs prompts on a cron schedule ("check if deploy finished every 5 min")

🤖 SUBAGENTS (Specialized Workers)

- Dedicated AI workers with their own context window, system prompt, and tool restrictions

- Route simple tasks to Haiku (cheap), complex ones to Opus (smart). Per-skill, per-agent.

- 5 built-in: Explore (read-only search), Plan (research), General-purpose, Bash, and Claude Code Guide

- Custom subagents defined in markdown — store at project or personal level

👥 AGENT TEAMS (Experimental)

- Multiple Claude Code instances working together with shared task lists

- Teammates message each other directly — not just reporting back to a main agent

- Best for: cross-layer changes (frontend + backend + tests), competing debugging hypotheses, research from multiple angles

⚡ HOOKS (17 Event Types)

- User-defined handlers that fire at lifecycle points: SessionStart, PreToolUse, PostToolUse, SubagentStart, PreCompact, etc.

- Three types: shell commands, HTTP endpoints, or LLM prompts

- Can block dangerous operations before they execute

📦 PLUGINS (Shareable Packages)

- Bundle skills, agents, hooks, MCP servers, and LSP servers into distributable packages

- Namespaced to prevent conflicts (/plugin-name:skill-name)

- Share your entire AI workflow as an installable package

Why this matters:

This isn't "AI writes code faster." This is a programmable agent platform where you control cost (model routing), safety (tool restrictions per agent), and orchestration (parallel subagents with shared context).

The shift from "Claude helps me code" to "I architect a team of agents that builds for me" just got a whole lot more real.

What's the first thing you'd build with this?

3 comments