@agenticgirl: This Anthropic talk made me realise Claude Code has a much bigger problem than writing code. Daisy Hollman, who works o…

This Anthropic talk made me realise Claude Code has a much bigger problem than writing code. Daisy Hollman, who works on Claude Code, put it simply: if Claude cannot do everything you can do, it cannot really do your job with you. Because your job is not sitting neatly inside the repo. It is spread across Slack threads, CI dashboards, design documents, internal tools, and years of knowledge that nobody bothered to write down. Claude Code starts with a repository and a shell. It has to be shown the rest. She also explained why every company needs to customize its coding agents. It is not because the model is not smart enough. Your team's way of working could never be fully included in its training because every codebase has different rules. Then came the part I had not thought about. Context windows have barely grown compared with everything agents can now do. So the difficult skill is no longer putting more information into the model. It is choosing what to leave out. Daisy compared it to running npm on an Arduino. On a laptop, you can keep adding things. Inside a limited context window, every piece of information is competing for space. Better feedback helps too. An agent that sees the equivalent of a red squiggly line immediately can correct itself before one mistake turns into ten. Daisy shared that Mozilla shipped more security fixes in one month with its newest model than it had in the previous 15 months combined. Claude helped create her slides and put the final point this way: "Your attention is the smallest box in the system." The problem is no longer getting enough information into an agent. It is helping one human keep up when 20 agents start sending work back.