We're All Addicted To Claude Code

Kelvin describes Codex’s early idea of agents working like coworkers who take a task and return with a pull request. He and the hosts discuss how Claude Code’s CLI encourages a different workflow from an IDE, including delegating codebase exploration to subagents with separate context windows.

The hosts and Kelvin argue that easy installation helps coding tools spread from individual engineers upward, while enterprise sales bring different security and control concerns. For building agents, Kelvin says context management is central: tools can use subagents, grep, and directory structure to find relevant code. He also says open-source code and documentation give agents material they can inspect and recommend.

Kelvin recommends reducing boilerplate, using well-structured packages, and giving agents tests, linting, and code review so they can check their work. He warns that agents tend to add code and can get stuck with poor context; he often clears context when it passes 50% usage. He says testing substantially sped up his work after he improved coverage.

Kelvin contrasts Claude Code’s human-like tool use with OpenAI’s emphasis on models that can handle longer tasks. He expects experienced engineers to benefit because they can direct work and judge architectural choices, while newer engineers still need to learn systems fundamentals. He also imagines smaller companies with personal cloud agents, and argues that product primitives and data systems will remain important as agents generate custom software.

Kelvin identifies context limits and agent orchestration as constraints, including how to verify merged code and bring in information from tools. He describes an agent finding a Rails job queue bug by tracing framework internals, and says Codex has caught complex behavior across multiple files that Opus missed. The conversation also covers prompt injection, sandboxing, and the tradeoffs of skipping permissions.