Understanding Bottleneck

A person still has to aim the work and judge the result after an agent has generated the options, the code, and the summary. What stays is the aiming — what is worth building, what matters, what is true enough, and how to steer.

Agents already generate options, write code, summarize sources, and produce alternative projections of information. The person still owns four things: what is worth building, what matters, what is true enough, and how to steer. Agentic Engineering is that work. This page is the human constraint on it. Context Engineering is how the person’s understanding gets into the agent’s window. Metacognition: The Control Layer is the steering itself.

A line that stuck in a 2026 conversation put it this way: you can outsource thinking but not understanding. The speaker endorsed a tweet he could not name. It is not his phrase. He also left a door open: maybe understanding gets automated too.

A personal wiki is one way to live under the constraint. Fixed data, many projections, so understanding forms in the person — and, in this vault, in the files. That file-half is ours, not his. The wiki should create better questions as well as answer existing ones. Useful outputs here: comparisons, contradiction reports, diagrams, alternate taxonomies, briefs, open-question lists, personal operating checklists. LLM Knowledge Systems owns the projections setup.

Aiming the work and judging the result stay with the person after the options, the code, and the summary have already moved. That is a constraint, not a method. The door stays ajar.

Related

• Metacognition: The Control Layer — the control layer that is the understanding the bottleneck names: steering, not producing
• LLM Knowledge Systems — the wiki-as-projections setup the conversation is describing
• Agentic Engineering — the work being directed; this page is the human constraint on that work
• Context Engineering — how the human’s understanding gets into the agent’s window

Open Questions

How can this wiki help the human understand instead of merely collect summaries?

What output formats best create insight — diagrams, questions, comparisons, or short briefs?

Sources

• Andrej Karpathy in conversation, Sequoia AI Ascent 2026. From Vibe Coding to Agentic Engineering. https://www.youtube.com/watch?v=96jN2OCOfLs Published 2026-04-29. The line about outsourcing thinking but not understanding is a tweet he endorsed and could not name; the bottleneck wording and the open door at 29:33 are his.