Mar 6, 2026 · 1h 11m · latent-space
Cursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth technical interview, Swyx speaks with Cursor's Jonas Nelle and Sam Whitmore about the launch of Cursor Cloud Agents, demonstrating how sandboxed Linux virtual machines, native computer use, automated video artifacts, and multi-model synthesis are redefining software engineering workflows.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.2 Guest teaching 3.8 Guest disagreement 1.1 The hosts pushing back 1.7
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 35:19 Pushing on the omission of a file editor
Swyx challenges Cursor's decision to drop the files app and compares it to OpenAI Codex's design smell, prompting Jonas to firmly defend forcing users to delegate to the agent.
Hardest push from the hosts ▶ 36:40 Challenging Cursor on deployment infrastructureSwyx pushes Jonas on whether Cursor should build an end-to-end hosting and deployment platform (Cursorapps) rather than leaving the loop open at code generation.
Biggest teaching moment ▶ 1:07:10 Reframing agent memory as self-awarenessSam educates Swyx on why static memory files fall short, outlining how dynamic file pointers and agent self-auditability of runtime constraints represent the true architecture for agent memory.
The host holds their own ▶ 40:08 Articulating the Agent Lab auto-router thesisSwyx demonstrates deep domain expertise by laying out his Agent Lab vs Model Lab thesis, explaining why agent labs must own model routing to abstract away provider loyalty.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Introducing Cursor Cloud Agents and the Autotab Integration | 3 | 4 | 1 | 1 | Swyx opens by asking whether the new Cloud Agents are essentially a repackaging of Autotab. Jonas explains the three core pillars: full VM test execution, automated demo video generation, and remote VNC control. | |
| Full-Stack Verification and the Brain in a Box Philosophy | 4 | 4 | 1 | 1 | Jonas demos full-stack verification where an agent writes a script in browser devtools to test size limits without explicit prompting. Swyx notes model capability milestones like Sonnet 3.5 enabling pixel automation. | |
| UI Polish, Multi-Model Comparison, and the Slash Repro Workflow | 5 | 3 | 1 | 1 | Swyx compliments the UI polish and connects slash repro workflows to classical ML concepts like reward hacking and test-driven development. Jonas explains using 20-second demo videos to evaluate Best-of-N model runs. | |
| Slash Commands, Datadog MCP, and Transcript Debugging | 6 | 4 | 1 | 2 | Sam explains how internal slash commands leverage Datadog MCP and transcripts for autonomous debugging. Swyx analyzes Datadog's strategic dilemma of whether to expose APIs via MCP or keep self-healing workflows proprietary. | |
| The Evolution of Coding: From Tab Autocomplete to Slack Workflows | 5 | 3 | 1 | 1 | Jonas and Sam describe how developer focus is shifting away from hand-coding and tab completion toward Slack-based agent orchestration. Swyx validates this pattern by sharing how non-technical teammates collaborate directly with coding agents in threads. | |
| Scaling Review Pipelines, Enterprise DevEx, and Team Configuration | 5 | 4 | 1 | 2 | Swyx brings up Graphite and debate over AI-assisted code reviews. Jonas explains that agent-driven code volume is forcing 10-person startups to adopt enterprise-scale deployment DevEx like stack diffs and merge queues. | |
| Virtual Machine Architecture, Snapshots, and Unshipped Features | 6 | 4 | 2 | 4 | Swyx compares stateful memory hydration against stateless Docker files and questions Cursor's decision to unship its file editor. Jonas defends the minimal UX philosophy of forcing users to delegate directly to agents. | |
| Deployment Platforms, Agent Labs vs. Model Labs, and Auto-Routing | 7 | 3 | 1 | 3 | Swyx pitches his Agent Lab vs Model Lab thesis and asks if Cursor should build a complete hosting platform like Vercel. Jonas explains why Cursor focuses on enterprise brownfield codebases rather than zero-to-one hosting. | |
| Parallel Multi-Model Execution, Model Synthesis, and Sub-Agents | 6 | 4 | 1 | 2 | Sam shares internal research on running an agentic synthesizer layer across diverse model providers. Swyx references Karpathy's council concept and probes how sub-agents are structured and routed in practice. | |
| Grind Mode, Inference Scaling, and Jevons Paradox in Software | 5 | 4 | 1 | 1 | Jonas describes long-running grind mode and Wilson's browser experiment, arguing that software velocity will scale through parallelism rather than raw model speed. Swyx draws parallels to parallel rollouts in reinforcement learning infra. | |
| Engineering Hiring in the Agent Era and High-Throughput Multitasking | 5 | 3 | 1 | 1 | The conversation shifts to engineering hiring in a token-rich environment and Jonas demonstrates rapid context-switching across agent tabs. Swyx discusses the impending convergence and conflict between coding tools and project management boards. | |
| Predictions, Dynamic File Memory, and Agent Self-Awareness | 5 | 6 | 1 | 1 | Sam reframes agent memory away from static rule files into dynamic file pointer contexts and harness self-auditability. Jonas and Sam discuss the frontier of agents becoming self-aware regarding environmental constraints and system prompts. |
Statements from this episode (24)
Nelle: No one had enabled AI coding agents to run code before Cursor
“Like obviously you need to run the code. And so that I think also is probably not that contrarian of a take, but no one has done that yet.”
Nelle: Full computer use shifted internal agent usage to driving new features
“Giving the model the tools to onboard itself and then use Full computer use end to end pixels in coordinates out and have sort of the cloud computer with different apps in it is the big unlock that we've seen internally in terms of usage of this going from, oh…”
Nelle: Video demos solve the code review bottleneck caused by autonomous agents
“We have found that in this new world where agents can end to end write much more code, reviewing the code is one of these new bottlenecks that, that crop up. And so reviewing a video is not a substitute for reviewing code, but it is an entry point that is much…”
Cursor disables Cloud Agents from spawning additional Cloud Agents
“We have disabled its cloud agents starting more cloud agents. So we currently disallow that.”
Nelle: Video Artifacts Make Best-of-N Multi-Model Evaluation Practical
“One of the things that's been a consequence of having these videos is we use best event where you run head to head different models on the same prompt. We use that a lot more because one of the Complications with doing that before was you'd run four models and…”
Nelle: Browser-Only AI Agents Fail When Workflows Require Native OS File Uploaders
“This is one of the cases where if you just do, you know, browser use type stuff, you will have a bad time because it now needs to upload files. Like it just uses its native file viewer to do that.”
Cursor launches BugBot autofix to automatically fix flagged issues
“We actually just launched BugBot autofix, where you can click a button and or change a setting, and it will automatically fix its own things.”
Cursor plans public release of transcript-based multi-agent debugging
“There'll be some versions of this as we ship publicly soon where you can spit up an agent and give it access to another agent's transcript to either basically debug something that happened.”
Cursor launches Model Context Protocol support for Cloud Agents
“We also launched MCPs along with along with this Cloud Agent launch, launch support for Cloud Agent MCPs.”
Nelle: Coding workflows will shift from inspecting diffs to previewing video demos
“That's going to happen again, where it goes from agents handing you back gifts and you're sort of like in the weeds and giving it, you know, 32nd to three minute tasks to you're giving it, you know, three minute to 30 minute to three hour tasks and you're gett…”
Nelle: Generating PRs is easy, but merge confidence is the new bottleneck
“One of the new bottlenecks is getting into production and we have a, Like, joke internally where you'll be talking about some feature and someone says, I have a PR for that. Which is like, it's so easy to get to, I have a PR for that, but it's kind of hard sti…”
Whitmore: AI Agents Will Eventually Automate Security and Performance Reviews
“There will probably eventually be things like performance level review, security review, things like that, where it's like more more different aspects of how this feature might affect your code base that you want to potentially leverage an agent to help with.”
Nelle: 10-Person Startups Using Cloud Agents Will Need Enterprise-Scale Dev Pipelines
“As with cloud agents, you scale up this parallelism and how much code you generate, 10 person startups become Need the DevEx and pipelines that a 10,000 person company used to need.”
Cursor uses Linux VM file system snapshotting to initialize agent environments
“The main default way is actually snapshotting. Like a VM, right? You run a bunch of install commands, and then you snapshot more or less the file system. And so that gets you set up for everything that you would want to Bring a new VM up from that template bas…”
Cursor removed web file editing to force users to delegate to agents
“And we actually felt that, that in some ways by restricting and limiting what you could do there, people would naturally leave more to the agent. And fall into this new pattern of delegating, which we thought was really valuable. And so there's currently no wa…”
Nelle: Developers do not hand-code or edit files directly anymore
“People like don't really edit files, hand code anymore. And so we want to build for where that's going and not where it's been.”
Whitmore: Synthesizing diffs across diverse LLM providers beats single-provider swarms
“What we found was that
At the time, at least there were strengths to using models from different model providers as the base level of this process.
like basically you could get almost like a synergistic output that was better than having like a very unified…”
Nelle: Sub-agents provide an architectural boundary to compress agent context cleanly
“And then the other reason to use sub-agents is we want, ah, contexts to be sort of summarized, reduced down at a sub-agent level.
That's a really neat boundary at which to compress that rollout and testing into a final message that that agent writes that then…”
Nelle: Major coding productivity unlock will come from agent parallelization
“The, we think that over the coming months, the big unlock is not going to be one person with a model getting more done, like the water flowing faster. It will be making the pipe much wider. And so parallelizing more, whether that's swarms of agents or parallel…”
Nelle: Cursor's agent swarms broke internal GitHub Actions CI/CD pipelines
“Like, we've broken our GitHub actions recently because we have so many agents, like, producing and pushing code that, like, CICD is just, like, overloaded because suddenly it's, like, effectively we grew, cursor's growing very quickly anyway, but you grow head…”
Nelle: Optimistic AI compute buildout forecasts still underestimate agent demand
“Even with You know, the most optimistic projections for what we're going to need in terms of build out are underestimating the extent to which these swarm systems can like churn at scale to produce code that is valuable to the economy”
Nelle: Developers will spend thousands to tens of thousands monthly on agents
“I think as we think about these highly parallel kind of agents running off for a long time in their own VM system, We are already at that point where people will be spending thousands of dollars a month per, per human, and I think potentially tens of thousands…”
Nelle: Cursor does not require latest AI coding proficiency when hiring
“I think that we don't see sort of necessarily being great at the latest thing with AI coding as a prerequisite.”
Nelle: Unguided AI models produce sloppy abstractions, requiring experienced engineers
“Models today do still have weaknesses, where if you let them run for too long without cleaning up and refactoring, the code will get kind of sloppy, and there'll be bad abstractions, and so you still do need humans that, like, have built systems before, know g…”