Building an Agentic Coding Editor (Cursor-style): The Complete Study Guide
Interview prep, subsystem by subsystem. ~77,000 words across 21 docs. Goal: read it aloud, internalize the mechanisms, and walk into a system-design / deep-dive interview able to explain how a Cursor-like editor actually works — and how you'd build one.
Everything resolves to three pressures: latency (typing must never wait on a model), economics (the median request must be cheap), and trust (a probabilistic actor with shell access must never silently wreck your repo). Every design choice below is a different answer to those three.
How to read this (start here) ¶
- Read
90-overview-architecture.mdfirst. It gives you the whole machine in one sitting: the architecture map, the keystroke-to-result flow for all 3 modes, and "the 10 ideas that explain everything." If you only have one hour, read this twice. - Then read the 16 subsystem sections (01–16) in order. Each is self-contained and ends with Interview Talking Points + Likely Interview Q&A (with model answers).
- Then the capstones:
91(design-it-from-scratch interview),93(mega Q&A bank),94(gaps + curveballs the sections don't cover). - The night before: read
92-cheatsheet-flashcards.mdonly. Glossary + 35 flashcards + numbers/names to drop.
File map ¶
The mental model ¶
90-overview-architecture.md— READ FIRST. Architecture map, 3-mode data flow, the 10 load-bearing ideas.
The 16 subsystems (each ends with talking points + Q&A) ¶
| # | File | What it covers |
|---|---|---|
| 01 | 01-editor-core.md |
Buffer data structures (rope/piece tree), render loop, VS Code multi-process model, fork-vs-extension |
| 02 | 02-language-intelligence.md |
LSP (semantic, project-wide) + Tree-sitter (syntactic, instant); diagnostics as agent feedback |
| 03 | 03-indexing-retrieval.md |
RAG for code: AST chunking, embeddings, HNSW, BM25 hybrid, Merkle sync, reranking, agentic search |
| 04 | 04-autocomplete-tab.md |
Fill-in-the-Middle, ghost text, contextual filter model, Cursor Tab = next-edit prediction |
| 05 | 05-agent-loop.md |
The while loop: ReAct think-act-observe, planning, guards, thin-harness vs scaffolding |
| 06 | 06-tool-use.md |
Function-calling contract, designing good tools, error-messages-as-teaching, parallel calls |
| 07 | 07-apply-model.md |
Turning intent into byte-precise edits: lazy sketch → fast merge model, search/replace, diffs |
| 08 | 08-context-management.md |
Prompt assembly order, lost-in-the-middle, KV/prompt caching, compression, stuff-vs-retrieve |
| 09 | 09-inference-infra.md |
Prefill vs decode, KV cache, PagedAttention, continuous batching, speculative decoding, routing |
| 10 | 10-diff-merge-vcs.md |
Myers/histogram diff, 3-way merge, checkpoints/undo, per-hunk review, git/PR, worktrees |
| 11 | 11-execution-sandbox.md |
PTY/terminal capture, allowlists/approval, containers/microVMs, background & remote agents |
| 12 | 12-mcp.md |
Model Context Protocol: hosts/clients/servers, transports, tools/resources/prompts, security |
| 13 | 13-safety-permissions.md |
Permission modes, prompt injection, the lethal trifecta, privacy mode, secret redaction |
| 14 | 14-evaluation.md |
SWE-bench (Verified), pass@k, acceptance rate, offline vs online, LLM-as-judge, regression suites |
| 15 | 15-model-training.md |
SFT, FIM objective, the apply model, RL-from-execution-feedback, distillation, the data flywheel |
| 16 | 16-system-design.md |
Backend view: client ↔︎ proxy ↔︎ providers/index/apply/exec; why a proxy exists even in privacy mode |
Capstones ¶
91-system-design-interview.md— "Design Cursor from scratch" 45-min walkthrough: clarifying Qs, requirements, components, deep dives, scaling, MVP→v2→v3, what strong vs weak answers sound like.92-cheatsheet-flashcards.md— Night-before cram sheet: 45+ term glossary, 35 flashcards, "numbers & names to drop," 20-item "if you remember nothing else."93-qa-bank.md— Mega Q&A bank: best questions from every section + 25 new cross-cutting ones, grouped by theme (incl. behavioral / "why-did-they-build-it-this-way").94-gaps-and-curveballs.md— Critic pass: accuracy corrections, coverage gaps with primers (monorepo scale, observability, on-device, cost modeling, eval contamination, agent memory, multimodal), and 15 curveball questions with answers.
Suggested 2–3 day plan ¶
Day 1 — Build the mental model + foundations.
90overview (twice). Then01core,02language-intel,03indexing,04autocomplete.- Test yourself: explain "what happens when I hit Tab" end-to-end out loud.
Day 2 — The agent half.
05agent loop,06tools,07apply,08context,09inference,12MCP.- Test yourself: explain "what happens when I tell the agent 'fix this failing test'" end-to-end.
Day 3 — Safety, eval, training, and the interview drill.
10diff/vcs,11exec/sandbox,13safety,14eval,15training,16system design.- Then
91design-from-scratch,93Q&A bank (answer out loud before reading the model answer),94curveballs. - Final pass:
92cheat sheet.
The 5 stories you must be able to tell cold ¶
- Keystroke → Tab completion. Debounce → FIM prompt (prefix/suffix + neighbor tabs) → tiny model + speculative decoding → contextual filter → ghost text → accept event feeds the flywheel. (
04,09) - Question → chat answer + edit. Retrieval (hybrid + rerank) → context assembly (stable prefix for cache, question last) → frontier model over SSE → apply model materializes the edit → diff review. (
03,08,07,10) - Goal → autonomous agent run. The
whileloop: system prompt + tools + history → tool call → harness executes (read/grep/edit/run) → observe (diagnostics, exit code) → self-correct → permission gate + checkpoint each step → bounded by budgets. (05,06,11,13) - Intent → byte-precise edit. Why not full-file rewrite or line-number diffs; lazy sketch + small fast merge model + unique-match search/replace + parse-validate + retry. (
07) - "Is it good?" → eval. SWE-bench Verified gates merges; acceptance rate + retention + A/B gate launches; the gap between "scores well" and "feels good" is the whole game. (
14)