Achieving the 'Zero Wait Thinking' Proven by Claude Code's Creator with DevLoop Runner
Boris Cherny, the creator of Claude Code, recently shared his development setup that has shocked the developer community. His approach involves running 10-15 Claude instances in parallel, switching between them using a notification system to ensure "human thinking never stops."
However, many developers who read about this face a common question:
"Is it even possible to replicate this individually...?"
This article explains the evolution of AI-driven development in three stages: how it used to work, the next-generation development style Boris proposes, and how DevLoop Runner makes it accessible to everyone.
1. Traditional AI-Driven Development
The Limitations of Sequential Processing
When AI-driven development first emerged, most developers used it like this:
Loading diagram...
This sequential processing style had several problems.
Wait Time: While the AI generates code, developers can only wait. If one task takes 3 minutes, five tasks result in 15 minutes of waiting.
Repeated Iterations: Using lightweight models leads to repeated cycles of implementation → review → fix. Even with "faster models," the increased number of iterations ultimately takes more time.
Manual Verification Overhead: Humans manually verify AI-generated code and request fixes if issues are found. This back-and-forth consumes time and drains developer focus.
Context Switching Cost: Even when trying to work on multiple tasks in parallel, it's difficult to track which task is in what state. High cognitive load reduces efficiency.
Why This Happened
In the early days of AI-driven development, the concept was to use "AI as a human assistant." Humans lead, AI assists. While this thinking isn't wrong, now that AI capabilities have dramatically improved, we need to redesign the development workflow itself.
2. The Next-Generation Development Style Boris Proposes
Boris Cherny redesigned the entire development workflow to break through these limitations.
Eliminating Wait Time Through Parallel Execution
Loading diagram...
Boris runs 5 Claude Code instances locally and an additional 5-10 in the browser in parallel. After submitting Task A, he immediately switches to another tab without waiting and submits Task B. He returns to tabs when notifications arrive—this cycle makes "wait time" essentially zero.
The "One-Shot Success" Strategy with Opus 4.5
Boris exclusively uses Opus 4.5 for all tasks. His conclusion: it's faster to get it right the first time with a heavy model than to iterate multiple times with lightweight models.
In his words: "Opus 4.5 is the biggest and slowest model, but it's almost always the fastest in the end because you have to steer it less."
Autonomous Verification Loop
Boris has set up an environment where "Claude verifies its own code." Using a Chrome extension to control the browser, he lets Claude autonomously iterate until "the UX feels good"—this is the secret to 2-3x quality improvement.
Instead of manual human verification, AI verifies itself, making verification wait time zero as well.
Leveraging Plan Mode → Auto-accept
Boris's workflow is as follows:
- Draft a plan in Plan Mode and have humans review it
- If the plan is good, switch to Auto-accept mode
- Claude completes the implementation in one go
Important decisions are made by humans, implementation is delegated to AI. This division of roles is clear.
Results: 259 PRs in 30 Days
With this workflow, Boris merged 259 PRs in 30 days. This translates to 497 commits, 40k lines added, and 38k lines removed. One person achieving the productivity of a small development team.
The Difficulty of Replication
However, replicating Boris's setup individually requires considerable time and expertise.
- iTerm2 notification setup and permission management
- MCP integration setup
- Multi-tab management system
- Git checkout management for each session
- Notification system optimization
And standardizing this across an entire team raises the bar even higher.
3. Achieving "Zero Wait Thinking" with DevLoop Runner
DevLoop Runner is a service that embodies the "zero wait thinking" Boris achieved through individual effort as a platform.
Actual Results
With DevLoop Runner, you can replicate Boris's level of productivity. In fact, development using DevLoop Runner commonly processes around 10 PRs per day, achieving productivity comparable to Boris's.
This isn't just a technical limitation but also the cognitive limit of humans. Reviewing and making decisions on 10-15 PRs per day is considered the practical upper limit for humans while maintaining quality. DevLoop Runner thoroughly eliminates "wait time" until you reach this limit.
No special setup or expertise required. The platform achieves "zero wait thinking" so you can focus solely on creative tasks.
Parallel Execution Managed Through UI
When you submit multiple Issues to Runner simultaneously, they run in parallel in independent environments. No need to manage multiple tabs like Boris. You can check overall progress at a glance in the UI, and all you need to do is think about "the next Issue."
No iTerm2 setup, no notification system construction needed. Open the dashboard in your browser and see all job states at a glance.
Automatic Model Selection
While Boris takes an "Opus 4.5 only" strategy, DevLoop Runner automatically selects the optimal model based on task characteristics.
High-performance models for phases requiring deep thinking, lightweight models for phases requiring fast processing—models are used appropriately for each situation. You can also combine multiple AI credentials (OpenAI, Claude, etc.).
You don't even need to make these decisions.
Autonomous Verification Loop
Loading diagram...
DevLoop Runner's 10-phase workflow automatically executes the loop: Issue → Code Generation → Test → Fix. Test scenario creation, test code implementation, and test execution are built into the standard process. If tests fail, it automatically returns to the fix phase and repeats until quality standards are met.
The "autonomous verification" Boris achieved with Chrome extensions is built in as a standard feature.
Balance of Human-in-the-Loop
You can get the best of both worlds from Boris's "Plan Mode → Auto-accept."
- Fully Automatic Mode: Simple tasks and routine work
- Staged Approval Mode: Important feature changes and modifications affecting architecture
You can flexibly set review points based on task characteristics. See the Execution Mode Guide for details.
Team-Wide Standardization
Turn Boris's individual technique into a standard environment for the entire team.
- No special setup required. Start within minutes of signing up
- Unify DevLoop Runner settings across the team
- New members can participate in "zero wait thinking" from day one
While Boris recommends each team maintain a CLAUDE.md file in their repository, you can achieve similar effects with DevLoop Runner by crafting your Issues well. See the GitHub Issue Writing Guide for reference.
For AI credential setup, see the AI Credentials Setup Guide. For team operations, GitHub Account Management is also an important point.
Getting Started: 3 Steps
Step 1: Experience Parallel Execution
Start by submitting 2-3 Issues simultaneously. Simple bug fixes, new feature implementations, refactoring—watching these progress in parallel, you'll experience the feeling of "waiting" disappear.
Getting Started with DevLoop Runner provides detailed instructions for your first job execution.
Step 2: Adjust Review Points
Flexibly set approval timing based on task characteristics. We recommend starting with staged approval mode to observe, then gradually switching to fully automatic mode for trusted tasks.
Step 3: Standardize Across Your Team
Once you experience the benefits, introduce it team-wide. The world Boris reached through individual effort can be shared by your entire team.
Summary: 15 Parallel Instances Are the Means, the Essence Is "Never Stopping Thinking"
What Boris Cherny proved is a new development style: "AI-era developers don't wait."
15 parallel instances is a shocking number, but it's just a means. The essence is creating an environment where human thinking is never blocked by I/O wait—and focusing that thinking solely on "the next creation."
Traditional AI-driven development was "waiting for AI" development. The next-generation development style Boris proposes is "making AI wait" development. Humans constantly keep thinking about the next task, AI keeps executing in parallel.
DevLoop Runner has platformized this essence. Without any special setup, you can get a "zero wait development environment" within minutes of signing up.
AI-era developers don't wait. They keep thinking.
Why not throw your next Issue to DevLoop Runner?
References
Get Started with DevLoop Runner
Auto-generate PRs from GitHub Issues. Let AI accelerate your development.