Agents

Skills give AI a set of instructions to follow — and instructions can be ignored. This essay maps Anthropic's dynamic workflows against a real writing checklist, tracing six structural gains from determinism to reproducibility.

AI engineering has evolved through three compensatory phases (prompt, context, and harness), each addressing a failure the previous layer couldn't fix. Harness Engineering is the governance layer that keeps teams of agents coherent across complex, long-running tasks.

This essay argues that AI is reshaping software at an architectural level, moving from human-centered applications to a composable agentic ecosystem where CLIs, Skills, and MCP form distinct layers that agents invoke as primary users.

Claude Code and CoWork share a codebase but diverge sharply — local vs managed-VM execution shapes resumability, attachability, and autonomy.

Claude Cowork marks Anthropic's pivot from chatbot to agentic AI: a productized Claude Code that plans, executes, and delegates tasks through sub-agents in a sandboxed Micro-VM environment.

This article traces the evolution of Claude's Skills, Commands, and Agents, analyzing the fundamental tension between intent-matching intelligence and explicit-command reliability, and arguing that their merger points toward compositional AI behavior.