Latent Space: The AI Engineer Podcast · Latent.Space

Cursor's Third Era: Cloud Agents

·1 hr 7 min·5 clips

Cursor's cloud agents now test their own code changes, running dev servers and iterating autonomously before delivering a ready-to-review PR.

Latent Space is a podcast for AI engineers and researchers. This episode is structured around a live product demonstration and technical explanation of Cursor's cloud agents launch, described as one of the biggest launches in the company's history. The hosts discuss an earlier experiment inside Cursor in which an LLM judge was not just selecting between model outputs but acting agentically — writing new code diffs by synthesising the strengths of multiple models from different providers, producing synergistic outputs better than either model alone. The main topic is cloud agents at cursor.com/agents. The core argument is that the key bottleneck for AI-written code is not the generation speed but the ability of the model to validate its own output. Prior cloud agents ran in blank VMs without proper development environment setup, essentially forcing the model to sight-read code and hope it was correct. The new system gives the agent a fully configured computer, lets it onboard itself to the repository, run development servers, and test changes end-to-end. The three pillars are described in detail. Pillar one: the agent tests its changes, running dev servers and iterating until it has a tested result rather than an untested I-tried-some-things PR. Pillar two: the agent produces a screen-recorded video of what it did, which serves as a much lower-effort entry point for code review than reading a raw diff — particularly valuable as diffs grow larger with agentic coding. Pillar three: the developer has full VNC remote desktop access to the agent's VM, allowing direct interaction, live previews, and precise diagnosis of what the agent produced. The hosts also discuss the future direction: the bottleneck is shifting from single-developer throughput to parallel agent fleets, and the pipe becoming wider rather than the water flowing faster.

As heard by us

Cursor's cloud agents are framed as parallel work for models and teams, with setup and VM sizing still evolving.

Cursor's cloud agents are cast as part of a broader shift toward agentic infrastructure: an LLM judge that can also write code, parallel agents, and blank VMs that run in their own computers.

Read the full review in PlayNext →

Why you'd press play

Cursor's cloud agents are about shared setup, parallel work, and the infrastructure choices behind both.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Latent Space: The AI Engineer Podcast