How Claude Code Claude Codes | The Vergecast

Anthropic's Claude Code, initially a developer tool, has evolved into a versatile AI agent capable of writing 100% of its creator's code and performing various non-coding tasks for a wide range of users, while raising significant data privacy and security concerns for AI tools in general.

• Claude Code, starting as a developer tool, now writes 100% of its creator's code and automates diverse tasks for all users.

• The product's evolution into Co-work highlights AI's immediate value in handling 'busy work' for non-technical users.

• Granting AI tools access to personal data requires extreme caution due to evolving policies and inherent privacy risks.

Claude Code, developed by Boris Churnney at Anthropic, has transitioned from assisting with 10% of coding tasks to handling 100% due to significant model improvements, particularly with Opus 4.5 and 4.6, which introduced advanced testing and self-correction capabilities. Although designed for developers, its broader utility led to the creation of Co-work, a more user-friendly version that excels at automating 'busy work' for non-technical users. However, the increasing capabilities of AI agents like Claude Code and Co-work, which require extensive access to personal data, necessitate a critical re-evaluation of data privacy, security, and the evolving relationship users have with these powerful tools.

Claude Code's Evolution

• 00:04:37 Claude Code, which launched a year ago, has become a pivotal AI product, evolving significantly in its capabilities. Initially, it handled only about 10% of its creator's coding tasks, but with the release of Opus 4.5 in November, its capability dramatically increased to 100%. This shift meant the model could not only write code but also test it and make fine-tuned adjustments, eliminating the need for manual human intervention.

The "Vibe Coding" Concept

• 00:04:52 The creator of Claude Code, Boris Churnney, no longer directly writes code, as Claude Code performs 100% of his coding. This change, while significant, felt like a natural evolution for an engineer accustomed to constantly adapting to new technologies. The fundamental shift is from engineers directly writing and scrutinizing source code to orchestrating AI agents that generate, test, and refine code, allowing engineers to focus on higher-level problem-solving.

User Base Expansion

• 00:10:55 While Claude Code was originally designed as a developer tool, it unexpectedly gained traction among non-developers, including data scientists, product managers, and even sales teams, due to its ability to automate tasks. This broader appeal led to the development of Co-work, a more accessible desktop application that offers features like deletion protection and a virtual machine, making it safer and easier for less technical users to interact with the powerful underlying AI agent.

AI for Busy Work

• 00:20:10 Co-work, built on the same Claude Code agent, has found immediate success by focusing on automating 'busy work' for general users. Examples include organizing screenshots, paying parking tickets, purchasing licenses, or handling project management tasks like sending Slack pings for status updates. This approach demonstrates that AI's initial widespread appeal might stem from solving small, tedious problems rather than attempting to build complex, all-encompassing life management tools.

AI Development Progression

• 00:29:16 The progression of AI capabilities is understood as moving from basic coding to tool use and then to full computer use. An 'agent' is defined as an LLM capable of using tools, which allows it to pull necessary context from large codebases or external systems without needing constant human input. The ultimate stage of computer use enables the AI to interact with any application, including complex websites lacking APIs, making the model incredibly versatile in automating diverse tasks.

Data Privacy & AI Risk

• 00:41:43 Interacting with AI tools like Claude Code and Co-work, which often request access to personal data like emails and calendars, poses significant privacy and security risks. Companies may change their data handling policies, and anonymized data can often be de-anonymized. Users should apply a 'assume it will be public' mindset and treat AI tools with a sharper eye than established services, especially free ones, as their data may be used for model training or subject to acquisition, potentially exposing sensitive information to unintended parties.