When should you use a tool vs. a skill vs. a subagent?

Anthropic's engineers frame it as choosing the right agentic primitive for the job — tools for actions, skills for reusable on-demand instructions, and subagents for isolated context — rather than piling everything into one orchestrator.

Why do AI agents get worse over time?

Because capabilities get bolted on without modernizing the architecture: the system prompt balloons, tools and subagents multiply, and the agent starts to regress where it used to excel.

Where can I read these Anthropic agent talks?

On PodLearn — full verbatim transcripts of Anthropic's Claude talks, searchable and free to read with an account.

Claude Agent Building — Anthropic Engineering Talks (Podcast Notes)

Episodes on PodLearn

Build a proactive agent workflow with Claude Code

Keep exploring

MCP explained by the people building it
Agent skills & progressive disclosure
Claude Code best practices, from Anthropic
Podcast MCP server — bring transcripts into Claude