Claude Agent SDK [Full Workshop] — Thariq Shihipar, Anthropic

I watched Thariq Shihipar’s workshop on the Claude Agent SDK and came away feeling both inspired and a little obsessed. If you build tools with AI, you’ll want to see this. Watch the full session here: Claude Agent SDK Workshop video.

At its simplest, the talk explains how an agent is more than a clever prompt, it’s an autonomous system that builds its own context and chooses its path. That distinction stuck with me, because I’d been treating models like one-off helpers, not self-managing collaborators. Anthropic’s SDK (formerly Claude Code) is designed to make those collaborations practical, solving recurring architectural headaches so you can iterate faster.

The workshop breaks an agent down into a practical “Harness” you can actually build right now: Tools, Prompts, a File System, Skills, Sub-agents, and Memory. One neat takeaway, and I say this as someone who’s debugged messy pipelines, is how powerful the Bash tool is. Letting an agent call existing software like ffmpeg or git feels like letting a junior dev use the right toolbox, instead of reinventing the wheel.

There’s also a tidy concept called *Context Engineering*, where files are both memory and a verification mechanism. I liked the suggestion to keep a CLAUDE.md file for self-documentation, it’s simple and brilliant. The agent loop Thariq demos follows three steps: *Gather Context, Take Action, Verify Work*. It’s iterative, and that verification bit helps the agent self-correct without retraining.

For bigger projects, he warns against brittle semantic search and recommends structured Claude MD files and scoping to subdirectories. He also shows how “Hooks” can inject deterministic rules when an agent drifts, which felt very practical.

If you build AI-driven workflows, this workshop gives you patterns you can adopt immediately. I’m already sketching ways to add a Claude-powered agent to a few projects, and you might too.