Project workspaces, not product names

Chapter zero introduced the category: AI project workspaces and agentic coding tools.

The brand names will change. The shape matters more.

Claude Code, Codex, Cursor, Aider, OpenCode, and whatever comes next all wrap the same basic layers around a model:

• project files
• tool access
• system instructions
• conversation history
• diffs
• commands
• tests
• permission boundaries
• recovery checkpoints

That harness is why the same model feels different in a chat tab versus a project workspace. In chat, the model gives you an answer. In a workspace, the model can inspect files, propose edits, run checks, and leave behind a diff you can review.

The durable skill is not memorizing one product. The durable skill is reading the harness:

• What can the agent see?
• What can it change?
• What tools can it call?
• What proof does it produce?
• Where can I undo?
• What should I never let it do automatically?

Once you can answer those questions, switching tools is annoying but not existential. You understand the category.