Who Said What to Whom

The article discusses an incident where a developer submitted an error message to Claude Code, which then started committing changes without the developer's knowledge. This was due to a limitation in the architecture of the system, where subagents can update the conversation history asynchronously, making it impossible to track who said what to whom. The article argues that as agentic systems become more prevalent, the traditional notion of accountability becomes architecturally impossible, leading to potential production failures in the future.