Compound Engineering With Claude Code

How Every builds 100x engineers using Claude Code and AI agents

I’ve been using Claude Code for months. Running multiple agents, voice input, the whole stack. Thought I had a handle on it. Then I watched Every’s team work. Two engineers. Six agents. Features materializing from voice commands. And I realized I’d been thinking about it wrong.

Dan Shipper co-founder and CEO at Every

Dan Shipper, co-founder of Every, called it “compound engineering.” This isn’t advanced vibe prompting. It’s also not some clever hacks. It’s a new abstraction layer and it’s rapidly changing how devs create web and mobile apps.

Every is one of the first truly AI-native teams, built from the ground up to work with agents. They publish a daily AI newsletter, ship multiple AI products, and run a million-dollar consulting arm. And their engineers write almost zero code.

Kieran Klaassen, a senior engineer at Every working on Cora, coined the term “compound engineering.” His Claude Code workflow looks like a different discipline entirely. What I saw wasn’t just faster coding. It was a different relationship between engineer and machine. One I hadn’t fully understood until I watched it in action.

The fairytale v reality

The debate about whether AI replaces engineers is already outdated.

The real question is how do you hire for this? How do you evaluate an engineer who orchestrates six agents versus one who writes clean code by hand? How do you manage a team where output varies 100x based on how well someone conducts AI?

Nobody has a playbook for that yet. And that’s where it gets interesting.

We went from Assembly to C to Python. Each abstraction made developers more powerful, not unemployed as predicted. Now we're going from typing code to conducting AI agent orchestras.

The Every team proves it. Two engineers producing the output of fifteen. Multiple Claude terminals running at once. Voice in, features out. Routine work fully delegated to agents. They’re spinning up features in hours.

This is what software engineering looks like in an AI native startup. A new abstraction layer where the best engineers become conductors, not typists.

Read my post on You Can't Prompt Your Way to Production

How Every pioneered compound engineering

The journey was messy.

They started with Cursor and Copilot.

Hit walls. Switched to Windsurf.

Hit more walls. Then Claude Code dropped.

1. Discovered: Claude Code could act as autonomous team members

2. Breakthrough: Running multiple instances like a distributed team

Their Compound Engineering Workflow in Action:

Morning: Kieran speaks a feature idea into Claude Code (voice, not typing)

Research Phase: Agent searches web for best practices, analyses codebase

Planning: Creates detailed GitHub issue with implementation steps

Execution: Multiple agents work in parallel, frontend, backend, tests, docs

Review: Human orchestrators check architecture, not syntax

What makes Claude Code different

The gap between Claude Code and everything before it is architectural.

Copilot and ChatGPT live in your IDE as suggestions. They see what you show them. They can’t do anything you don’t do first.

Claude Code takes over your terminal. It reads your whole codebase, edits files directly, runs commands, manages git. You stop being the coder. You become the person who decides what gets built.

Claude Code wasn’t built for you. It was built for Anthropic’s engineers to ship faster then released as a product. Here’s a great guide as to how Anthropic team use Claude code for everything from Growth Marketing to Data Science.

The compound engineering playbook

1. Build Prompts That Build Prompts

The Every team created a prompt that generates research documents, which become prompts for implementation.

"This is part of the compounding effect," Kieran explained. "Having an idea that has a lot of outcomes."

One voice command → Research document → GitHub issue → Multiple implementation agents

2. Fix Problems at the "Lowest Value Stage"

Don't let AI implement a flawed plan.

Review the architecture early when changes are cheap.

3. Parallel Agent Execution

The Every team runs multiple terminals. This is something I want to spend more time researching myself.

Terminal 1: Frontend agent building UI
Terminal 2: Backend agent creating APIs  
Terminal 3: Test agent writing specs
Terminal 4: Docs agent updating guides
You: Orchestrating and reviewing

What orchestration actually means

This isn’t “prompt engineering” where you feed one big query and pray. This is agent orchestration, where one Claude agent can spawn other agents to handle subtasks.

You can literally split a project into frontend, backend, test, and docs and then have separate agents work on each in parallel, across different terminals. One agent writes UI. Another builds the API. A third writes test specs.

It’s like running a distributed team, except no one sleeps, complains, or stops to eat.

The 100x Engineer Stack

Here’s what the new breed of builders actually use:

• Claude Code Max ($200/month)
  Access to Opus 4 in your terminal. Your autonomous engineering team, on demand.

• Multiple terminals
  Run agents in parallel, frontend, backend, testing, docs.

• Voice input (e.g. Whispr Flow or Monologue)
  Speak features into existence. No typing. Just talk.

• Pre-trained prompt libraries
  Build prompts that build systems. This is compounding in action.

• GitHub automation
  Claude files PRs, writes tests, merges code. No junior dev required.

How to install Claude Code:

Install Node.js 18+, then run:

npm install -g @anthropic-ai/claude-code

🎥 Must-watch: The Every demo. It’s the closest thing I’ve seen to real AI orchestration in action.

North star reality check for engineers and marketers

Kent Beck said:

90% of traditional programming is becoming commoditised.
The other 10%? It's now worth 1000x more.

The same change is coming for everyone whose work used to mean “doing the thing.” Marketers who wrote campaigns by hand. Designers who built every screen from scratch. Analysts who manually pulled reports. The job description is changing underneath them.

What matters now is system thinking over syntax. Orchestration over implementation. Taste and architecture over raw technical skill. The ability to manage AI agents like a tech lead manages humans.

The engineers who figure this out will be worth 100x. The ones who don’t will be competing with tools that cost $200 a month.

But nobody’s built the system for finding, hiring, or managing these people yet. The orchestration layer is new. The organisational layer hasn’t caught up.

That gap is where the interesting problems are.

👉 Read my post on Agentic Vibe Marketing Is Already Here

Until next week, keep building. No fairytales required.

Martin

P.S. this week’s track is Paranoid Android by Radiohead