Why Most Tools Built Specifically for AI Agents Will Fail

Summary

The article argues that most attempts to build software specifically for AI agents will fail because current agents remain broadly human-like in how they interact with computers and information. Tools that work well for agents therefore tend to work well for humans too, while existing products benefit from being represented in training data. A new tool would need to overcome the advantage agents gain from already knowing established software, programming languages, libraries, patterns, and idioms. The article is particularly skeptical of creating a new programming language for agents, since agents may be less effective in an unfamiliar ecosystem even if it offers some technical improvements. It also argues that the ideal ergonomics of agent tools remain unclear: claims about preferences such as static typing are plausible in opposite directions, and practical measurements have not established a clear answer. Those assumptions may become obsolete as model capabilities and context management change; the article notes that compaction has recently become much more capable, including repeated re-compaction of a 272k-token context window. The author recommends incremental improvements such as plain-text or Markdown access, functional APIs, MCP servers, and CLIs. For now, “building for AI agents” largely means prioritizing the API over the UI, but the article questions whether that distinction will last as computer-use systems improve, citing GPT-6-Astra as an example.