Cursor's Third Era: Cloud Agents
Cursor's Third Era: Cloud Agents
Cursor's Third Era: Cloud Agents
This episode dives into Cursor's latest cloud agents launch—a major evolution in AI-powered software development that moves far beyond autocomplete and local coding assistants.
Cursor's cloud agents grant AI full computer control within isolated VMs, enabling end-to-end feature development, automated testing, and video-based output review—shifting developers from line-by-line editing to high-level oversight. Key innovations include three-pillar workflows (test, demo video, remote VM access), slash commands like `/repro` for bug reproduction and `/council` for multi-model orchestration, and subagents for context-aware task delegation. The platform emphasizes parallelism over latency, using swarms and best-of-N evaluation across diverse models to boost throughput. Cursor deliberately minimizes web UI surfaces—including removing the file editor—to encourage agent delegation, while prioritizing team workflows via Slack integration, MCP plugins, and persistent VMs. Though not building full-stack hosting, Cursor focuses on bridging the gap between PR drafts and production-ready code through enhanced review layers, memory-aware onboarding, and self-auditing agents—signaling a fundamental shift toward AI-native development.
00:00
00:00
Using models from different providers had strengths, resulting in a synergistic output
00:53
00:53
Cloud agents now give the model full computer use, enabling driving new features
02:08
02:08
The agent returns a tested PR ready for review, similar to what a human should do when asking for a review
02:57
02:57
Testing is enabled by default for PRs, but can be skipped with /no test or customized in agents.md
03:35
03:35
Demo videos are an easy entry point for code review and help with alignment, especially when requests are underspecified
06:10
06:10
Around Claude 4.5, full pixel-based automation became feasible
10:49
10:49
/repro automatically reproduces bugs, records videos of the process, and fixes them
32:02
32:02
The web UI aims to keep it minimal and expose shared app surfaces for both users and agents
32:39
32:39
Removing the Files app encourages delegation to agents instead of manual file handling
34:34
34:34
The big bottleneck is getting code out, not deployment infrastructure
36:34
36:34
Every agent lab should have a router to decide the best model for users
40:54
40:54
Andrej Karpathy called it the council, and there's an internal command '/council' written by Ian
1:03:35
1:03:35
Agents must be self-auditable and self-aware to understand their environment, secrets, and system constraints
