Build with Multiple AI Agents using Claude Code in 40 Minutes

Klaassen contrasts rapid, back-and-forth vibe coding with handing an agent a larger task and reviewing its work later. He says agents work best when given clear directions about the problem and how to approach it, and that deciding what to build still requires product thinking.

Klaassen shows Claude Code exploring a Rails 8 demo app and using to-do lists to track work. For an email summarization feature, he asks it to research database models, a summarization service using RubyLLM, and an email UI, then document the plans as separate issues. He says parallel subagents can run several tasks while keeping their detailed work out of the main conversation context.

Klaassen uses Git worktrees to give separate Claude Code sessions their own branches and folders, reducing conflicts when agents edit code at the same time. He notes that tasks may depend on one another, so the plan should account for shared assumptions and dependencies. In the demo, agents add code, migrations, and tests, though the app also hits problems with an outdated library version.

Klaassen creates pull requests and uses review commands and subagents to check changes from different perspectives, including security and business concerns. He says research and code review can be bigger bottlenecks than writing code. He also turns repeatable workflows into slash commands, such as researching issues, proofreading, and reviewing pull request comments.

Klaassen describes Kora, which sorts Gmail messages by importance, archives less urgent email, and summarizes it for users. He says AI workflows should learn from user preferences and that repeated work can be captured in commands so it becomes easier the next time. His advice to beginners is to try one simple task and build from there.