Cursor¶
AI-native code editor and the company behind it. Started in 2022–23 as "spicy autocomplete" (levels-of-autonomy-shapiro level 1), evolved to tab-completion, to chat, to agent, to multi-agent orchestration.
Cursor 3 (April 2026)¶
Complete rewrite — no VS Code base anymore. Streamlined for agent-first workflow. First tab of the product is multi-agent orchestration: nested agents, project-level status aggregation, explicit "here's what you as a human need to review" surfaces.
Motivation per Zakariasson: "we are using more agents and we need a better control panel" — the traditional IDE layout doesn't match the work anymore.
Cloud Agents¶
Each agent gets its own VM. Agent can:
- Start dev environment via an internal cursor dev tool backend start / cursor dev tool frontend start CLI that abstracts orbstack, clickhouse, postgres, redis, electron, etc.
- Run the actual Cursor app against its own repo clone.
- Use computer control to test its own work — agent records a video of itself clicking through the UI and returns it for review.
"Multiple thousands of agents per day" running against copies of the Cursor codebase internally.
Rules, skills, hooks¶
- Cursor Rules — "most misunderstood concept we launched." Emergent rather than pre-installed: make rules when agents go off the rails, not because a directory says to. See emergent-cursor-rules.
- Skills + MCPs — give agents capabilities (e.g. add-a-feature-flag skill, so autonomous PRs self-flag their changes).
- Hooks — block agents from touching sensitive code paths (auth, encryption).
Bugbot¶
Integrated PR reviewer — looks at GitHub diffs and comments automatically. One leg of the "plan → produce → review" SDLC automation stack.
Cross-references¶
- eric-zakariasson — internal engineer/dogfooder
- software-factory — his framing of what Cursor enables
- claude-code — adjacent tool; Hashimoto's choice in the zed-editor context
- agent-as-junior-engineer — same philosophical frame