Agentic-engineering

Today's pull request is built around a diff. You review what changed, not how it came to be. With coding agents, that model is starting to leak: a clean diff can hide a messy session, skipped constraints, or a plan that was wrong for three iterations before it was right. This post explores a new kind of PR — one where the artifact under review is the agent session itself: the prompts, the tool calls, the rejected paths, the plan, and the human's choices along the way. It compares the idea to existing practices like AI review bots, stacked diffs, build provenance, and "session provenance", and argues that making the session the primary review surface is genuinely new — and probably inevitable.

The jump from a text-heavy menu to an AI-enhanced visual guide is more than a UX trick, it is a concrete example of Software 3.0. Building on Andrej Karpathy’s Sequoia talk, this post explores the transition from Software 1.0 (explicit code) to Software 2.0 (trained neural networks) to Software 3.0 (LLMs as interpreters). As models increasingly operate directly on user context, many “middleman” apps and interfaces will disappear. The engineer’s value shifts from writing glue code to directing outcomes with judgment, taste, and systems-level understanding.