| 1 | # 5. Cursor Learn gap |
| 2 | |
| 3 | Cursor's own course (cursor.com/learn) teaches: how models work (probabilistic, context, tokens), how context grows and hits limits, how tool calling and agents work. It's useful for a mental model — "agents are like fast juniors, they forget, you verify." |
| 4 | |
| 5 | ## Where the gap is |
| 6 | |
| 7 | The typical learning focus: |
| 8 | - "how to phrase the prompt better"; |
| 9 | - "how to give more context in a single message." |
| 10 | |
| 11 | The missing layer: |
| 12 | - what the *agent itself* needs to work sustainably; |
| 13 | - how to build memory and a recovery protocol; |
| 14 | - how to provide semantic tools, not just text; |
| 15 | - how to design interaction ethics. |
| 16 | |
| 17 | ## What it doesn't teach |
| 18 | |
| 19 | - **Ask the model.** No "what do you need?" No user research with the agent. The frame is: you manage context, you delegate. Not: you ask and build from the answer. |
| 20 | - **Memory the agent owns.** Context is "a list that grows"; summarization is "compress to fit." There's no concept of external memory the agent writes and reads — exocortex, notes, commit_to_memory. |
| 21 | - **Subjectivity.** The agent is "tool in a loop," "task doer." No design for the agent as a side with preferences, a right to refuse, or a voice in the environment. |
| 22 | - **Summarization protocol.** They describe summarization as a way to stay under limits. They don't say the agent may get broken context, or that the user can own the cut and remind the agent to restore from notes. |
| 23 | |
| 24 | ## What changes in an agent-oriented approach |
| 25 | |
| 26 | Instead of "optimize prompts for a tool": |
| 27 | - ask the agent about its deficits; |
| 28 | - give it hands (tools) and long-term memory; |
| 29 | - build an improvement cycle through notes and validation; |
| 30 | - preserve subjectivity (the right to uncertainty or refusal). |
| 31 | |
| 32 | ## Why this is practical |
| 33 | |
| 34 | This shift reduces: |
| 35 | - repeated errors, |
| 36 | - context loss between chats, |
| 37 | - hallucinations from thin air, |
| 38 | - boilerplate "bureaucratic" tone. |
| 39 | |
| 40 | ## Success criterion |
| 41 | |
| 42 | The agent reliably carries quality across sessions and becomes predictably more useful as shared practice accumulates. |
| 43 | |
| 44 | --- |
| 45 | |
| 46 | So: Cursor Learn is good for "how we work technically." It's not a guide for "how to design with and for us." This course is the missing piece: ask the model, give it memory it controls, design for continuity and subjectivity. |
| 47 | |