Creating and Managing Multiple AI Agents Revolutionizes Programming

Boris Cherny, the creator of Claude Code at Anthropic, shared his innovative programming approach that has caught the attention of the software engineering community.

In a discussion on X, he revealed how he uses five AI agents to optimize his workflow, turning programming into a real-time strategy game experience.

"They call this best practices, and if you’re not following them, you’re falling behind as a programmer," stated Jeff Tang, an influential member of the developer community.

Working with Multiple Agents

The core of Cherny’s new approach is the management of five Claude agents in parallel. He uses iTerm2 to organize tasks, where each tab in the terminal corresponds to a different agent.

"I run 5 Claudes in parallel on my terminal. I numbered my tabs from 1 to 5 and use system notifications to know when a Claude needs input,” Cherny explained.

This way, one agent can run tests while another refactors existing code, increasing the efficiency of the development process and allowing a single developer to achieve the output of a small team.

Slower Model, Greater Accuracy

Cherny revealed he uses Anthropic's more robust model, Opus 4.5, despite it being the slowest. "It's the best coding model I've ever used. Even though it's larger and slower, it ends up being faster in the end," he said.

This insight highlights a crucial aspect: prioritizing a more capable model reduces the time spent on debugging since the model is more accurate from the start.

Collective Memory and Learning from Mistakes

Cherny’s team maintains a file called CLAUDE.md, where they document the mistakes made by the agents. This helps avoid repetitions and continuously improves the model.

"When a developer reviews a pull request and finds a mistake, the AI is tagged to update its instructions. Each mistake becomes a rule," explained Aakash Gupta.

Automating Repetitive Tasks

Cherny’s work is propelled by rigorous automation of recurring tasks. He uses custom bar commands to perform complex operations quickly.

An example is the command /commit-push-pr, which automates the bureaucracy of version control, allowing Cherny to focus on more critical tasks.

The Edge of Code Verification

The verification of implemented changes is a differentiator of Claude Code, where the AI not only generates text but also performs testing. "Claude checks all the changes I make using the Chrome extension," Cherny states.

By allowing the AI to verify its own work, the quality of deliveries increases significantly, with Cherny noting that this improves the final outcome by up to three times.

Implications for the Future of Software Engineering

Cherny's approach suggests a shift in how developers view their work, emphasizing that AI can be a workforce rather than just a mere assistant.

Developers who adapt to this new mindset will not only become more productive but also have the chance to completely transform their programming methods.