
Where does control really live in agentic AI systems? The harness.
Sufian Kaki Aslam is a Senior Cloud Architect who builds production AI systems for enterprise clients and teaches in our AI Engineering with Claude Nanodegree program. His recent blog breaks down why engineers are increasingly building their own harnesses.
To preview some key insights:
✔️ Every Claude product you use — Claude for Chrome, Claude Code, Claude Cowork — is a harness Anthropic built. They're general-purpose, which means they're not optimized for your specific use case.
✔️ When you build your own with the Claude Agent SDK, you define the rules: which tools the agent can access, how it manages context, when it needs human approval before acting.
✔️ A useful way to think about API vs. SDK: contractor vs. employee. The API handles one-time tasks. The SDK is for agents with ongoing responsibilities that need to behave reliably at scale.
✔️ The learning curve isn't syntax, it's judgment: knowing when to constrain an agent, how to write guardrails that actually hold, and how to structure a loop that fails gracefully.
Full breakdown here: udacity.com/blog/what-is-t…
English














