Jagged Intelligence: What to Delegate — Agent Engineering

Agent Engineering · Unit 2 of 6

Activation

Reflection

Concept

Practice

⚡ Activation

💭 in your head · nothing to type

2024 example: Claude/GPT could refactor 100,000 lines of code — and miscount the number of "r"s in "strawberry"; newer models pass this example.

Don't picture this as "AI with bugs" — picture it as a savant. Someone who is uncannily brilliant in one dimension and utterly helpless at something a child handles in a second.

Run two moments in your head: one when AI floored you with its capability — and one when AI floored you with its cluelessness. Both moments are about the same system. What could possibly make both true at once?

Guess before the explanation

Getting it wrong is fine: it is part of learning. Guessing first helps the answer stick later, and the answer itself comes after the explanation.

A pipeline has the step “count the exact number of items in an order.” By the logic of jagged intelligence, how should you handle it?Give the model a tool (calculator, DB) or make this step codeHand it to the model: it copes with 100,000 lines of code, so it will cope with countingAsk the model to count twice and take the averageRemove the step from the pipeline entirely

Learn with AI

Copy your personal prompt and paste it into your agent's learn mode — it picks up your context and takes you forward.

Honestly: with AI it feels like you can do more than you can without it. After practice, check yourself without AI. (Bastani et al., PNAS 2025)

Explains the material using the course method.

ChatGPT →Claude →Gemini →Copilot →