Full Tutorial: Build with Multiple AI Agents using Claude Code in 40 Minutes | Kieran Klaassen

Summary

Kieran Klaassen describes a shift from “vibe coding”—working closely with AI in a rapid feedback loop—to agentic coding, where you delegate larger, well-defined tasks to AI agents and review their results.

He frames the developer’s role more like a manager’s: decide what problem to solve, provide clear instructions, coordinate work, and check the output. This approach can also support research, product planning, and writing—not just coding.

Key workflow and techniques

• Research before implementation: Ask Claude Code to inspect the repository, explain how it works, and research a proposed feature. Klaassen recommends planning first rather than immediately asking the tool to code.
• Write down the plan: Turn research into Markdown specifications or GitHub issues. These preserve the rationale and make tasks easier for teammates or other AI tools to pick up.
• Delegate independent tasks in parallel: Subagents can work on separate parts of a feature simultaneously. This can save time and keep task details from consuming the main conversation’s context window. Tasks should be sufficiently independent, with dependencies considered.
• Use Git worktrees for parallel coding: Each agent can work in a separate branch and folder, reducing the risk of multiple agents editing the same files. Worktrees require care when switching between them.
• Scale reasoning to the task: Klaassen uses prompts such as “think carefully” for complex or production-sensitive work, but says simple edits do not need extensive deliberation.
• Keep human review in the loop: Agents can write code, run tests, and handle database migrations, but they may still make mistakes—especially with unfamiliar or changing frameworks. Pull requests and direct inspection remain important.
• Automate recurring work with slash commands: Custom commands can package repeatable instructions for tasks such as repository research, issue creation, code review, security checks, or applying a preferred writing style. Klaassen recommends turning repeated tasks into reusable workflows.
• Review from multiple perspectives: Separate review agents can take on roles such as security researcher or entrepreneur. Other bots or models can add different perspectives. AI reviews may catch issues, but their findings still need to be verified.

Claude Code and tool choice

The demo uses Claude Code from a terminal, though Klaassen emphasizes that the workflow matters more than the specific tool. He contrasts its focused, text-based interface with more visual coding environments such as Cursor, while encouraging developers to experiment as tools and personal preferences change.

He also demonstrates voice-to-text input and shortcuts to reduce friction. A mode that skips permission checks can be convenient but potentially risky, so he recommends caution and appropriate isolation rather than using it blindly.

Live feature-building demo

The pair plan a simplified email product involving stored emails, summaries, and a user interface. Klaassen first asks the tool to research the project and break the work into tasks, then launches implementation in separate worktrees. The demo illustrates the process but does not show every feature fully finished or thoroughly reviewed.

Cora email assistant

Klaassen presents Cora as an email assistant designed to make an inbox more manageable. It categorizes and prioritizes messages, surfaces important items, summarizes newsletters and other mail, and can archive less urgent messages for later review.

Users can correct classifications, allowing the assistant to learn their preferences. It can also draft emails. Klaassen says the system combines multiple language-model workflows and rules because people differ in what they consider important and how much risk they are comfortable with.

Advice for newcomers

Start with one small, practical task rather than trying to automate an entire workflow. Experiment, inspect what the AI does, and develop a process that suits your needs instead of treating any one person’s method as universal. A starter project can also give the AI a consistent technology stack and established components to build on.

Main speakers and sources

• Kieran Klaassen — Cora founder and engineer; demonstrates the agentic-coding workflow and Cora.
• Peter Yang — interviewer and host.