Anthropic just gave away its agent-building playbook
Anthropic dropped the full material from its Code with Claude workshops onto GitHub, and it's the closest thing to an official manual for building real agents. Nine modules, TypeScript and Python, Apache 2.0, github.com/anthropics/cwc-workshops. Not a toy, this is the stuff they teach engineers in the room.
What's inside tells you what Anthropic thinks matters. Picking the right model for quality-per-dollar. Decomposing one big agent into a multi-agent system with Skills and MCP. Giving agents memory that survives across sessions. Eval-driven development, meaning you don't ship an agent until a test harness says it got better. And two full production builds: an M&A research team called Deal Desk, and a SEC-filings research desk that dispatches analyst sub-agents.
The interesting signal isn't any single module, it's the shape. Every serious pattern this year, Skills, MCP servers, persistent memory, eval loops, multi-agent decomposition, shows up here as the default, not the frontier. Anthropic is quietly standardizing how agents get built and handing you the reference implementation for free.
If you're building on the Claude Managed Agents API, this is the first thing to read. Repo: github.com/anthropics/cwc-workshops
← Back to all articles
What's inside tells you what Anthropic thinks matters. Picking the right model for quality-per-dollar. Decomposing one big agent into a multi-agent system with Skills and MCP. Giving agents memory that survives across sessions. Eval-driven development, meaning you don't ship an agent until a test harness says it got better. And two full production builds: an M&A research team called Deal Desk, and a SEC-filings research desk that dispatches analyst sub-agents.
The interesting signal isn't any single module, it's the shape. Every serious pattern this year, Skills, MCP servers, persistent memory, eval loops, multi-agent decomposition, shows up here as the default, not the frontier. Anthropic is quietly standardizing how agents get built and handing you the reference implementation for free.
If you're building on the Claude Managed Agents API, this is the first thing to read. Repo: github.com/anthropics/cwc-workshops
Comments