Latent Space Podcast: Cursor's Third Era – Cloud Agents

Date: March 6, 2026
Podcast: Latent Space: The AI Engineer Podcast
Episode: Cursor's Third Era: Cloud Agents
Host: Latent.Space
Guests: Key Cursor team members – “A”, “B”, “C” (from transcript; names not provided but representing founders and engineers at Cursor)

Overview

This episode explores the launch and implications of Cursor’s new Cloud Agents, seen as ushering in a "Third Era" for the company. The discussion traverses the evolution from code autocompletion to truly agentic, autonomous software engineers in the cloud. The team explains core features, design decisions, UX shifts, technical challenges, the future of collaborative engineering—and how agentic coding is changing the industry.

Key Discussion Points & Insights

1. Evolution to Cloud Agents

• From Local to Cloud-Based Agentic Workflows:

  • Initially, models acted as assistants for small, local tasks (autocomplete, tab completion).
  • The new “Cloud Agents” paradigm allows models to operate on dedicated cloud computers with broader system access, setting the stage for more autonomy and impactful work.
  • B [01:09]: “Giving the model the tools to onboard itself and then use full computer… have a cloud computer with different apps… is the big unlock that we've seen internally… going from, oh, we use it for little copy changes to… really driving new features with this kind of new type of agentic workflow.”

• Key Features Launched:

  • Automated Testing: Agents now test their own changes end-to-end before requesting code review, imitating best engineering practices.
  • Demo Videos: Agents submit short videos showing what was accomplished, providing faster alignment and context for reviewers.
  • Remote VM Access: Users have full remote control over agent VMs—terminal, desktop, etc.—enabling hands-on review and further experimentation.

• Parallelism and Swarms:

  • The vision for “wider pipes, not just faster water”—meaning real productivity leaps will come from running many autonomous agents in parallel (swarming) rather than a single, ever-more-capable model.
  • B [00:30]: “Over the months the big unlock… will be making the pipe much wider and so paralyzing more. Whether that's swarms of agents or parallel agents…”

2. Product Demo & UX Philosophy

• Testing & Videos as First-Class Artifacts

  • Testing is now the default (with opt-out options). Demo videos present agent output, lowering review barriers and aligning expectations earlier.
  • A [05:24]: “The videos have been super helpful… Having the video up front makes that alignment… a shared artifact with the agent very clear, which has been just super helpful for me.”
  • Videos commonly include UI walkthroughs, screen captures, and, when relevant, dev tools interactions.
  • B [03:20]: “Reviewing a video is not a substitute for reviewing code, but it is an entry point that is much, much easier to start with…”

• Full VM Access & Live Interaction:

  • Agents run on persistent VMs that simulate a real developer workspace (including desktop, terminal, secrets management, etc.).
  • Users can “jump in”—watch demos, modify outputs, or directly iterate within the environment.

• UX Details & Branding:

  • Cursor’s design choices, such as branded cursors, thematic wallpapers, and integrated video “Chapters,” enhance both usability and product delight.
  • C [09:09]: “Obviously a company called Cursor has to have a beautiful cursor… And then when you download it, it adds a little cursor kind of TikTok clip thing.”

3. Agent Workflows & Parallelization

• PR Generation & Review Bottlenecks:

  • Automated agents can generate, test, and produce PRs for complex features and bugfixes, shifting the human role to higher-order decisions and UX calls.
  • B [10:02]: “We use Best of N… you run head to head different models on the same prompt… But if you come back with four 20 second videos… you can figure out like, which one of those do you want to iterate with to get it over the line?”

• Slash Commands Ecosystem:

  • Rich internal command system (e.g., /no test, /repro, /diagnosis) powers agent orchestration, reproducing bugs, debugging, and more.
  • A [14:53]: “One of my favorites is Cloud Agent Diagnosis… spin up a bunch of subagents using the Datadog MCP to explore the logs and find all the problems…”

• Transcript Sharing & Agent Forking:

  • Agents can transfer transcripts, act as external debuggers, or continue conversations from other agents’ sessions, adding collaborative intelligence layers.

4. Cloud Agent Technical Decisions & Tradeoffs

• Infrastructure Choices:

  • Various VM/image options: Docker-based setup, full VM snapshotting, support for persistent desktops.
  • Partial unshipping of experimental features, such as in-browser file editing, in favor of patterns that encourage agent-based work over manual intervention.
  • B [33:54]: “You want to be able to work at really high levels of abstraction and double click and see the lowest level. That's important. But… in some cases limiting the UX capabilities makes for a cleaner experience…”

• Persistent, Stateful Agents:

  • Persistent VMs allow agents (and users) to leave and resume work, preserving context and system state beyond stateless Docker runs.

• Model Selection & Routing:

  • Users can select models for agents; Cursor plans to enhance “auto” routing for best-in-class choices as model capabilities evolve.
  • C [37:47]: “Put me on cursor auto. You decide for me based on all the other people—you know better than me.”

5. Agent Swarms, Subagents, and the “Society of Agents” Model

• Best-of-N and Model Swarms:

  • Multiple models can simultaneously act on prompts; outputs can be synthesized by meta-agents, sometimes achieving “synergistic” results.
  • A [41:00]: “There were strengths to using models from different model providers… you could get almost like a synergistic output that was better than having a very unified… bottom model tier.”

• Subagents:

  • Internally and externally, subagents are spun up for specialized sub-tasks (exploration, computer use, etc.), supporting parallel progress on complex objectives.
  • B [43:31]: “We wanted to do a lot more to expose hooks and ways for people to configure these… another example of a built-in subagent is the computer use subagent in the cloud agents…”

• Long-Running and Recursive Agent Work:

  • The team recounts building full browsers and multi-day agent grinds, emphasizing planning/validation steps before unleashing large-scale autonomous work.
  • A [46:21]: “We really mean three days… There’s been a long time with the grinds.”

6. Collaboration & Team Development Patterns

• From Solo to Collaborative Dev via Slack:

  • Engineering work starts to take place in communication tools (Slack) where issue threads, agent invocations, and PRs are all visible and annotatable by various team members.
  • Enables distributed, asynchronous, and more collaborative product development.
  • A [20:44]: “For us, Slack is almost like a development IDE basically… It turns into almost like a discussion service where people can… collaborate on UI… and you can end up with a PR again with the artifacts visible…”

• Role Specialization Changes:

  • Humans focus on high-level product, UX, and architectural decisions. Agents fill in implementation detail and repetitive engineering labor.
  • Agents are now responsible for validation (with tools like Bugbot) before handoff.
  • B [21:31]: “The nugget of what is actually interesting and relevant… is not the boring details of where does this if statement go. It's: do we want to ship this? Is this the right UX?… leave the implementation to the cloud agent.”

7. Scaling, Infra, and The Road to Production

• Emergent Bottlenecks:

  • Bottlenecks move "up the stack": once code generation and PR creation are automated, merging, review, and deployment become constraints.
  • The need for mature CI/CD infrastructure is shifting from large enterprises to every team—even small startups.

• Production-Ready Agents:

  • Ensuring that autonomous PRs are tested, readable, and robust enough for production is still a challenge—requiring both automated and human review.
  • Bugbot and similar tools are increasingly relied on for code-level guarantees.

• PLG & Team Onboarding:

  • Product-led growth patterns: individuals adopt Cursor for side projects, bring it into team work, and then teams require advanced features (marketplaces, shared config, etc.).
  • B [27:14]: “We started working on… making it really great for teams and making it that the 10th person that starts using Cursor… is immediately set up…”

8. The Future: AI Agents & The Changing Developer Landscape

• Shifts in Dev Identity:

  • The team muses on the obsolescence of manual (“hand”) coding, agent layers vs. “model labs,” and the future of product-embedded, multi-modal AI.
  • B [18:58]: “Hand coding… So Boomer.”

• Cost and Productivity Leverage:

  • Expectation: token/inference costs per developer could enter the thousands or tens of thousands per month. But such costs are justified by huge leverage.
  • B [51:03]: “Individuals get that much more leverage and if one person can do as much as 10 people… that tool… is going to be tremendously valuable and worth investing in…”

• Hiring for the New Era:

  • The fundamentals—engineering sense, ability to reason about abstractions—remain critical, but agent proficiency is now an accelerant.
  • A [52:34]: “Cursor also operates very quickly, and leveraging agentic engineering is probably one reason why that's possible in this current moment.”

• Predictions:

  • Transition from local to cloud agent dominance will complete within the year.
  • “Zero lines of code” (manual) per developer, with all output generated by agents in the near future.

Notable Quotes & Moments (with Timestamps)

• Parallelism over Model Speed:
  “[00:30] B: …the big unlock is… making the pipe much wider… whether that’s swarms of agents or parallel agents, both of those… contribute to getting much more done in the same amount of time.”

• Testing & Demo Videos:
  “[05:24] A: …having the video up front just has that makes that alignment… a shared artifact with the agent very clear, which has been just super helpful for me.”

• Bugbot Reliance:
  “[24:44] B: Once that happens two or three times, you learn to wait for Bugbot.”

• Agentic Development Patterns:
  “[20:44] A: …Slack is where a lot of development happens. Like we will have these issue channels or just like this product discussion channels where people are always at cursoring and that kicks off a cloud agent.”

• Design Philosophy:
  “[33:54] B: …you want to be able to work at really high levels of abstraction and double click and see the lowest level… But… in some cases limiting the UX capabilities makes for a cleaner experience…”

• Model Synergy:
  “[41:00] A: …you could get almost like a synergistic output that was better than having a very unified… bottom model tier. So it was really interesting…”

• The End of Manual Coding:
  “[18:56] C: Oh, that's cute. [about ‘hand coding’]”
  “[58:02] A: …probably zero lines of code written in the whole month of December this year. By myself, 100% agents is a personal prediction…”

Additional Highlights & Innovations

• Cloud Agent Diagnosis:
  Uses Datadog MCP to automatically launch subagents for deep debugging, slashing troubleshooting time.
• Transcript Forking:
  Passes session context between agents for enhanced collaboration and error resolution.
• Persistence:
  VMs and agent sessions are persistent, supporting long-running or paused/resumed work.
• Marketplace/Plugin Model:
  Shared skills and MCPs for teams, admin-manageable configuration.
• Slash Command Power-Users:
  Internal commands facilitate custom workflows (e.g., autorepair, test skipping, reproduction).

Timestamps for Major Segments

• Cloud Agents Introduction & Purpose – [00:00]–[03:00]
• Product Demo (Testing, Videos, VM Remote Control) – [02:07]–[06:00]
• Design Details & Team Collaboration – [09:06]–[12:00]
• Command System & Advanced Operations – [13:35]–[17:00]
• Evolution of Agentic Coding – [18:00]–[22:00]
• Scaling/PLG for Teams & Marketplace – [26:24]–[28:20]
• Technical Infrastructure Choices – [28:24]–[34:00]
• Parallelism, Subagents, Model Swarms – [39:20]–[43:50]
• Deep Dive on Long-Running Agents & Society Model – [45:12]–[48:20]
• Operational Bottlenecks, Hiring in Agentic Era – [51:03]–[53:03]
• Predictions, Wrap-up Thoughts – [58:02]–[64:00]

Tone and Language

The tone is highly technical but pragmatic, with a playful, collaborative spirit. Cursor’s team displays humility about tradeoffs, excitement about user experience, candidness about technical debt and product decisions, and continuous curiosity towards emerging patterns in AI engineering.

Summary

Cursor’s Cloud Agents mark a milestone in the evolution of developer tooling—moving from assistive, code-generating LLMs to fully autonomous, agent-driven software engineering on persistent cloud infrastructure. The team describes a future built on parallelism (“swarming” agents), best-of-N model orchestration, agent-driven testing and review, and new collaborative patterns that blur the lines between traditional IDEs and communication tools like Slack.

Manual “hand coding” is fading, and the bar for productivity is now defined by how well you orchestrate, instruct, and review agentic workflows. As these systems gain greater autonomy and continue to scale, they’re reshaping what it means to create software, organize teams, and even hire. Cursor charts a course not just for their product, but for the next generation of software engineering—faster, more collaborative, and powered by fleets of tireless cloud agents.

For more technical deep-dives, visit latent.space