Claude Agent SDK reveals five-layer kit

Anthropic’s Claude Agent SDK is starting to look less like a vague “agent framework” and more like a real kit with distinct layers. The pieces were rolled out across several releases, but together they form a pretty clear architecture: persistent instruction files, reusable skills, event hooks, delegated subagents, and plugin bundles. That matters because the hard part of building agents is usually not raw model intelligence. It’s getting memory, tool use, orchestration, and guardrails to stop fighting each other. (anthropic.com) ### What is the SDK, exactly? Basically, it’s the same underlying harness that powers Claude Code, exposed so other teams can build their own agentic experiences. Anthropic has been pretty direct about that. The SDK gives developers access to the same context management, core tools, and permissions framework Claude Code uses internally, and it has already been pitched for things like financial compliance, cybersecurity, and debugging agents. (an([anthropic.com)# Why are people calling it a five-layer kit? Because the docs and product materials now point to five recurring building blocks that do different jobs. CLAUDE.md handles standing instructions and project memory. Skills package domain know-how into structured folders. Hooks fire actions at specific moments in the agent lifecycle. Subagents let one agent delegate work to another. Plugins bundle reusable components like slash commands, agents, MC(anthropic.com)on a single page — it’s an inference from Anthropic’s own docs and launch posts — but it’s a useful mental model. (anthropic.com) ### What does CLAUDE.md actually do? Think of CLAUDE.md as a README written for the agent instead of the human. Anthropic’s materials describe it as the place for project structure, commands, style rules, and other persistent instructions. In larger repos, Claude can discover multiple CLAUDE.md files up the directory tree, which lets teams scope guidance by repo, package, or folder. So this layer is less “memory” in the chatbot sense and more durable operating context. (resources.anthropic.com) ### Why are Skills a separate layer? Because instructions alone are too blunt. Skills are structured directories with a SKILL.md file plus scripts and resources, so an agent can load specialized know-how only when it needs it. Anthropic’s key idea is progressive disclosure — preload just the name and description, then pull the full skill into context when the task calls for it. That keeps context lighter and makes expertise portable across workflows. (anthropic.com) ### Where do Hooks and Subagents fit? Hooks are the automation glue. They let developers trigger shell commands or checks at specific points — run tests after edits, lint before commits, send notifications, log behavior. Subagents are the parallelism layer. Anthropic describes them as specialized workers that can take on delegated tasks while the main agent keeps moving, which is a big deal for longer, multi-step jobs. (anthropic.c([anthropic.com)are the packaging layer. Anthropic says Claude Code plugins can bundle slash commands, agents, MCP servers, and hooks behind one install step. That turns a pile of custom wiring into something a team can actually share and reuse. It also hints at where this is going — from handcrafted agent setups toward something closer to an ecosystem. (anthropic.com)Why does this matter beyond coding? Because Anthropic is steadily moving these pieces into broader products. Xcode’s native integration with the Claude Agent SDK brings subagents, background tasks, and plugins into Apple’s IDE, which shows the SDK is not just a terminal toy. The pattern is becoming clear: one harness, many surfaces, and more room for teams to plug in internal tools over MCP without rebuilding the whole stack each time. (anthropic.com) ### Bottom line? The interesting part is not that Anthropic has five named features. It’s that those features now separate the agent’s “brain” from its memory, procedures, triggers, helpers, and tool bundles. That makes the SDK easier to reason about — and much easier to integrate into real workflows where one giant prompt was never going to be enough. (anthropic.com)