Anthropic's developers share tips, playbooks and opinions on building with Claude and Claude Code, including agents, skills, context engineering and prompt caching.
Why I recommend it: Written by the company that makes Claude, so it's practical but not neutral. Best read by people who already use Claude for coding.
A long explainer on what an "agent harness" is — the loop, tools, memory, sandbox, permissions and budget caps wrapped around an AI model — its six core components, and why the wrapper now matters more than which model you pick. Opens with an agent that burned $400 in API calls overnight in a retry loop.
Why I recommend it: Free to read, no sign-up, and the best single walkthrough of the six pieces if you are trying to understand why agents that demo well fall over in real use. Two honest caveats: it is published by a consultancy that sells help with exactly this, so the framing favors building a harness; and the byline on it is not a person I can verify anywhere independently, which on this site means treat the article as a useful explainer rather than a sourced authority. For neutral definitions, the Wikipedia entry on agent harness and Addy Osmani's write-up cover the same ground.
