Claude Code’s Secret Workflow: How 5 AI Agents Turn Coding Into a Real‑Time Strategy Game

Claude Code’s Secret Workflow

When Boris Cherny, the creator of the world’s most advanced coding agent Claude Code, spilled the beans about his terminal setup, the whole Silicon Valley stopped typing and started commanding. What looks like a handful of tabs in iTerm2 is actually a battle‑ready fleet of five autonomous Claude agents that can write, test, refactor, document and even ship code – all in parallel.

---

The Problem: Traditional Linear Development Is a Bottleneck

• Single‑threaded inner loop – write a function → run tests → fix bugs → repeat.

• Human correction tax – every AI‑generated line still needs a developer’s eye, turning speed gains into wasted minutes.

Developers have been stuck in this “type‑then‑wait” rhythm for decades, and the productivity ceiling is hitting the wall.

---

The Solution: Parallel Claude Agents & Smart Orchestration

1. Five Agents, One Terminal

“I run 5 Claudes in parallel in my terminal, number my tabs 1‑5, and use system notifications to know when a Claude needs input.” – Boris Cherny

Using iTerm2 notifications, each tab runs a dedicated workflow:

• Tab 1 – runs the full test suite.
• Tab 2 – refactors legacy modules.
• Tab 3 – drafts API documentation.
• Tab 4 – generates new feature code.
• Tab 5 – performs a final verification run.

The result? A real‑time‑strategy‑style development environment where the programmer is the commander, not the sole soldier.

2. The Slowest, Smartest Model – Opus 4.5

Instead of chasing the fastest token‑generation, Cherny relies on Anthropic’s Opus 4.5, the heaviest model in the lineup. It may be slower per token, but because it makes fewer mistakes and excels at tool‑use, the overall time‑to‑completion is 2‑3× faster.

“You have to steer it less and it’s better at tool use, so it’s almost always faster than using a smaller model in the end.” – Cherny

3. One Shared Knowledge File – CLAUDE.md

AI agents suffer from “amnesia”. Cherny’s team solves it by keeping a single markdown file in the repo that records every mistake the agents make. When an error is caught, the team adds a rule to CLAUDE.md. The next time Claude runs, it reads the rule and avoids the same pitfall.

This turns the codebase into a self‑correcting organism.

4. Slash Commands & Sub‑Agents

Custom slash commands (e.g., /commit-push-pr) let a single keystroke trigger a full commit‑push‑pull‑request cycle. Specialized sub‑agents handle:

• Code‑simplifier – cleans up architecture after the main work.
• Verify‑app – runs end‑to‑end UI tests via a Chrome extension.

5. Verification Loops Are the Real Unlock

Claude doesn’t just write code; it tests it. Using the Claude Chrome extension, the agent opens a browser, runs UI tests, and iterates until the UI feels right. This verification loop is what many insiders credit for Claude Code’s rapid $1 billion ARR.

---

Who Benefits?

| Audience | Why It Matters | |----------|----------------| | Developers | Multiply personal output by 5× without hiring extra engineers. | | Team Leads / CTOs | Reduce “correction tax” and accelerate delivery pipelines. | | Product Managers | Faster feature validation, lower risk of regression bugs. | | AI Researchers | A real‑world case study of agent orchestration and model‑selection trade‑offs. |

---

What This Signals for the Future of Software Engineering

• AI as a workforce, not just an autocomplete.
• Parallel agent orchestration will become a standard pattern – think of it as DevOps meets RTS gaming.
• Model choice will shift from “fastest” to “smartest” as the human‑in‑the‑loop cost dominates.
• Self‑learning knowledge bases (like CLAUDE.md) will evolve into company‑wide AI policy files.

If you’re still typing line‑by‑line, you’re playing a single‑player mode while the rest of the industry is moving to multiplayer.

---

Dive Deeper

• For a deeper look at how agentic workflows are reshaping development, check out the article Awesome Agentic Workflows.
• Want to see the exact setup that powers five parallel Claudes? Read Claude Code Creator Workflow: 5 AI Agents Parallel Coding Real‑Time Strategy Game.
• Learn how Anthropic’s Co‑Work AI Agent File Management keeps knowledge alive across sessions in Anthropic Co‑Work AI Agent File Management.

---

Keep Up with the AI‑Powered Development Revolution

For more cutting‑edge analysis, follow Agent Arena – the hub where engineers, product leaders, and AI enthusiasts share the latest breakthroughs.

---

External references:

• Real‑time strategy games (Wikipedia)

• Large language models (Wikipedia)

• Anthropic Opus 4.5 announcement