How would you measure the success of the Agents Window launched in Cursor 3?

Introduction

Cursor 3, released on 2 April 2026 under the internal codename "Glass", added the Agents Window: a full-screen workspace that sits next to the classic editor and lets developers start, watch and manage many agents at once, whether they run locally, in separate git worktrees, over SSH or in Cursor's cloud. It is Cursor's clearest bet that the job of a developer is shifting from typing code to directing agents.

Measuring its success is tricky. More agents running is not automatically good: agents can burn credits, produce code nobody merges, or overwhelm reviewers. I'll define success in terms of useful work shipped, with guardrails on quality and cost.

Step 1

Product Context (3 minutes)

The Agents Window is the control centre for parallel agent work in Cursor. Users can:

• Launch several agents on different tasks, each in its own worktree or cloud machine.
• See status for every agent: planning, running, waiting for input, finished, failed.
• Compare outputs when they run the same task with different models (for example Composer 2 versus a frontier model).
• Review diffs, accept or discard results, and move work into a pull request.

Key stakeholders:

• Individual developers on Pro, Pro+ and Ultra plans who run several tasks at once.
• Engineering teams on Teams and Enterprise plans who care about throughput and code quality.
• Cursor, which earns more when usage grows on paid plans but bears model costs, especially on its own Composer models.
• Reviewers and maintainers, who receive the output of these agents.

User flow:

1. Open the Agents Window and describe a task.
2. Choose where it runs (local worktree or cloud) and which model (or Auto).
3. Agent plans and executes; user optionally steers.
4. User reviews the diff and test results.
5. User keeps, merges or discards the work.

Lifecycle stage: early growth. The surface is new, power users are adopting quickly, and mainstream users are still learning how to work with several agents at once.