CyberAgent's ¥300M AI Experiment: How 1,200 Engineers Transformed Software Development with Claude Code and Codex

In the rapidly evolving landscape of software engineering, generative AI has transitioned from a novel assistant to an active agent. Leading Japanese megaventure CyberAgent recently made headlines by investing approximately 300 million yen (USD $2 million+) annually to integrate cutting-edge AI coding tools into its development workflows.

Targeting over 1,200 engineers, this massive corporate initiative deployed tools like Anthropic's Claude Code and GitHub Copilot (powered by OpenAI's Codex/GPT models). One year into this ambitious rollout, CyberAgent has shared invaluable insights into how these AI agents have altered daily development, which tools succeeded, and what limitations still exist.

The Core Initiative: Bridging the Gap Between Developers and AI Agents

CyberAgent’s strategic goal was simple yet profound: maximize developer velocity by shifting from simple autocomplete tools to autonomous AI agents.

Historically, tools like GitHub Copilot acted as "copilots," suggesting line-by-line code completions. However, with the advent of agentic tools like Claude Code, AI can now understand entire codebases, run terminal commands, execute tests, and self-correct errors based on error outputs.

Main Entities Involved:

• Implementing Organization: CyberAgent, Inc. (a major Japanese internet services and media company).
• Primary AI Products:
  • Claude Code: An agentic, command-line interface (CLI) tool developed by Anthropic, built on the Claude 3.5 Sonnet model.
  • GitHub Copilot / Codex: The widely-used developer assistant ecosystem powered by OpenAI's specialized LLMs.

Key Findings: What Worked and What Didn't After One Year

After deploying these tools at a massive scale, CyberAgent observed several distinct shifts in developer productivity and sentiment.

1. The Rise of the "Agentic" Workflow

Traditional AI tools required developers to copy-paste code snippets into a browser interface. With Claude Code, engineers can issue high-level commands directly in their terminal, such as:

"Find the bug in the authentication middleware, fix it, and make sure all tests pass."

The AI agent scans the repository, identifies the target files, applies the edits, runs npm test or equivalent commands, and presents a completed git diff to the engineer. This has significantly reduced context-switching.

2. High Adoption vs. Tool Fatigue

While the initial excitement led to a surge in usage across various experimental models, the engineering team eventually consolidated around two primary platforms:

• Inline completion (GitHub Copilot): Remains indispensable for low-latency, real-time typing assistance.
• Autonomous agent tasks (Claude Code): Preferred for complex debugging, refactoring legacy codebases, and writing comprehensive test suites.

3. The "Unused" Tools

Generic chat interfaces that lacked repository context or terminal access saw a sharp decline in usage among developers. If an AI tool requires too much manual hand-holding (e.g., manually uploading files), developers naturally abandon it in favor of integrated CLI or IDE agents.

Comparing the Leading AI Coding Solutions

To understand how these tools fit into CyberAgent’s workflow, let's compare the primary technologies utilized during the trial:

Challenges and Lessons Learned

Despite the overwhelming success, CyberAgent encountered several notable bottlenecks:

• Token Consumption and Costs: Agentic AI tools that repeatedly scan large codebases consume high volumes of tokens. CyberAgent's 300 million yen budget highlights that enterprise-level AI agent deployment requires careful financial monitoring.
• Hallucinations in Complex Logic: While Claude Code is highly capable, it still occasionally invents dependencies or misunderstands deep business-specific domain logic. Code reviews remain non-negotiable.
• Security & Compliance: Ensuring proprietary codebases are not used for public model training was a critical prerequisite. CyberAgent utilized enterprise-grade API agreements to guarantee data privacy.

Frequently Asked Questions (FAQ)

Q1: Is CyberAgent replacing human engineers with AI?

A: No. CyberAgent's initiative is designed to augment human engineers, not replace them. The AI agents handle repetitive tasks, boilerplate code, and initial debugging, allowing human developers to focus on architecture, system design, and product requirements.

Q2: What is the main difference between GitHub Copilot and Claude Code?

A: GitHub Copilot excels at real-time, inline code completions as you type. Claude Code acts as an agent; it can interact with your computer's terminal, read files, edit code across multiple files, and run tests autonomously to verify its work.

Q3: How did CyberAgent manage the security of their source code?

A: CyberAgent established secure enterprise pipelines and API contracts with Anthropic and GitHub/OpenAI, ensuring that none of the code processed by the AI agents is used for training public models.

Conclusion: The Future of Developer Experience (DevEx)

CyberAgent’s one-year retrospective proves that AI coding agents are no longer a futuristic concept—they are a core component of modern enterprise development. By investing heavily in tools like Claude Code and GitHub Copilot, they have successfully shifted the developer's role from "writer of code" to "editor and director of AI agents."

As these models become faster and more cost-effective, organizations that embrace agentic workflows will likely outpace those relying solely on traditional development practices.