Claude Code agent teams pass tests without reading code

The article is a practical guide for people who already know Claude Code basics — asking, permission modes, CLAUDE.md, /clear and /rewind — and want to hand over more work. It groups five features from videos #1–#5 into two families: writing down procedures and rules (skills, hooks, permissions), and splitting up and delegating work (subagents, agent teams). Each was tested in a small practice repository, with results checked on screen.

Within the delegation family, subagents and agent teams serve different needs. Subagents work in their own context and return only a summary; the article's measurement showed that handing over a 22-file research task cut what stayed in the conversation. Agent teams go further: the first conversation becomes a coordinator, launches roles, and lets those roles talk directly. The splitBill example showed the test agent writing tests from specs alone, never reading the implementation, and passing. But the article notes teams carry coordination overhead and use significantly more tokens, so the official guidance is to check whether a lighter method suffices first.

The stakes here likely hinge on how much coordination between roles a task actually needs. If direct back-and-forth between builder and tester is not required, subagents appear to deliver the benefit at lower cost. For teams already automating coding and testing with Claude Code, the practical question may be where the token budget pays off — and whether an experimental feature becomes a standard one.