Cursor's Last Six Months - From Editor to Agent Workspace
Canonical version: Cursor's Last Six Months - From Editor to Agent Workspace.
Six months ago Cursor was an AI code editor. Over four releases since then, it added its own models, parallel agents and a window to manage them, and the editor is now one part of an agent workspace.
Cursor became a model lab
Anysphere started shipping its own frontier models. Composer 1 arrived with Cursor 2.0 as a fast, mixture-of-experts coding model. Composer 2 followed in March 2026. Then Composer 2.5 landed in May, matching Claude Opus 4.7 and GPT-5.5 on coding benchmarks at well under a dollar per task.
All three are built on Moonshot AI's open-weight Kimi K2.5 checkpoint, so Cursor spends its compute on post-training and reinforcement learning instead of building a base model from scratch.
Four kinds of agents
Cursor also added several ways to run agents:
- Cursor Subagents (v2.4, January) split a task into parallel pieces, each with its own context.
- Cursor Automations (March) run always-on, triggered by a schedule, a Slack message, or a GitHub event.
- Cursor Cloud Agents graduated to general availability and can now be self-hosted.
- Cursor Multi-Agents run up to eight agents at once, each in its own git worktree.
Cursor 3 made agents the interface
Cursor 3 shipped in April 2026. It added the Agents Window, a workspace that shows every agent you are running, local, in a worktree, in the cloud, or over SSH, all in one place. New commands came with it. /worktree isolates a task in its own working copy. /best-of-n runs the same task across several models in parallel and compares the results.
The Agents Window is optional. If you only want Cursor Tab completions and chat, the old editor is still right there.
Customization caught up
Cursor also added ways to control what the agents do. Cursor Skills (v2.4) package domain knowledge into SKILL.md files the agent loads only when they are relevant. Cursor Hooks run scripts at fixed points in the agent loop, so you can block an unsafe tool call instead of hoping a rule discourages it. Cursor Blame labels which lines were written by AI and which by a human.
What I make of it
Cursor is betting that the unit of work is moving from "a developer editing files" to "a developer supervising agents." Composer 2.5 lowers the cost of running many agents, the Agents Window puts them in one place, and skills, hooks and Blame give you some control over what they do.
I think the bet is right. The risk I see is that supervising eight agents is a skill of its own, and the tooling assumes you already have it.
Related
- Cursor
- Anysphere
- Composer 2.5
- Composer 2
- Cursor Agent Mode
- Cursor Multi-Agents
- Cursor Skills
- Cursor Hooks
- Cursor Blame
- Cursor Automations
- Cursor Cloud Agents

About Sébastien
Ready to get to the next level?
Found this valuable? Share it with someone who needs it.
