The AI Coding Shakeout Is Here

SpaceX absorbs Cursor. Roo Code shuts down. What happens next.

Two major pieces of AI coding news dropped this week. One involves a $60 billion acquisition option, the other a shutdown notice, and both stem from the same market dynamics.

On Tuesday, SpaceX announced a deal with Cursor that includes an option to acquire the company for $60 billion by year-end, or pay $10 billion for “work together.” Cursor gets access to xAI’s Colossus supercomputer — reportedly equivalent to a million H100s — to train their next-generation Composer model.

The same day, Roo Code announced it’s sunsetting its VS Code extension, cloud service, and router on May 15. The company is pivoting to Roomote, a Slack-based cloud agent. Their CEO Matt Rubens was direct about the reasoning: “We don’t believe IDEs are the future of coding.“

Their VS Code extension—three million installs, and a tool that pushed the boundaries of autonomous agents—is being archived.

Both events trace back to the same market pressure.

Changes to the AI Harness World

Foundation models are getting better and cheaper, and the gap between what Claude and GPT can do—and what their open-weight competitors can do—narrows with each release. Anthropic and OpenAI are both shipping their own coding tools now, competing directly with products that use their APIs. If your product is “only” one tool—an IDE extension, a CLI, a developer-facing wrapper—differentiation gets hard fast.

Cursor’s situation was awkward. As TechCrunch noted: “Cursor still uses and sells access to Claude and GPT models even as both firms roll out their own coding tools.” You can build a great product on top of models you don’t control, but you’re always one upstream decision away from having the rug pulled. When your suppliers become your competitors, the economics stop working in your favor.

Cursor was too close to the frontier labs to avoid becoming one of them. The SpaceX deal is that transformation made explicit: access to a million H100s to train your own models, absorption into the SpaceX/xAI ecosystem, and a path from wrapper to infrastructure provider. As a Dev.to analysis put it: “If you do not control the layer beneath you, your growth is conditional.”

We Still Love IDEs, But Not Everyone Does

Roo Code made the opposite choice. Instead of going deeper into the stack, they’re getting out of the IDE game entirely. Rubens described their internal team already moving to cloud environments where agents run multiple tasks in parallel without direct oversight. The IDE becomes a place you visit to review work, not where work happens.

It’s a coherent position, but we disagree. Every independent developer, every engineering team, every enterprise shipping production software still lives in an editor for most of their working hours. That isn’t changing anytime soon.

The real question isn’t IDE vs. cloud — it’s where value accumulates. Cursor’s answer is to control compute and own the model. Roo’s answer is to escape the editor and build autonomous agents that deliver finished work. Both are valid responses, but they share an assumption: that the IDE is either a shell for proprietary AI or a legacy surface to escape.

An agentic engineering platform rejects that framing. Instead of owning the model or abandoning the editor, it integrates developers’ existing tools — letting users move tasks between mobile, VS Code, JetBrains, or the CLI with persistent session state and context. Instead of proprietary models, it orchestrates parallel workflows by routing to the most suitable of 500+ models for each task. Teams can standardize best practices into reusable modes and track ROI from a centralized dashboard. The platform layer doesn’t need to own the compute or bypass the editor — it needs to make every developer faster regardless of which model or interface they prefer.

What We’re Building at Kilo

We’ve been watching this consolidation happen from our corner of the market.

We shipped a blog post this week about the Roo shutdown titled “Thank you, Roo.” Kilo started as a fork of Roo—a lot of what we do well today started with work they shipped first. Custom modes, the Architect/Code/Debug split, the whole “let the agent actually do things” philosophy.

However, we disagree that the IDE is over. Independent developers, engineering teams, enterprises shipping production software still live in an editor for most of their working hours. That isn’t changing anytime soon, and the more important variable is model lock-in, not IDE vs. cloud.

We just rebuilt the Kilo VS Code extension based on some of these same industry-wide changes that have been occurring, and the need for:

• True parallel execution.

• Subagent delegation.

• An Agent Manager for running and monitoring multiple agents at once.

• Cross-platform sessions that carry state between terminal and editor.

The underlying philosophy: model-agnostic, open source, focused on engineering workflows. We’re not building a Claude wrapper vulnerable to API changes, or a walled garden training proprietary models on a million GPUs.

What This Means for Developers

Takeaways from this week:

Dependency is risk. Cursor built a great product, but they were dependent on models they didn’t control. The SpaceX deal is expensive insurance against that dependency. If you’re betting your workflow on a tool, understand who controls the layers beneath it.

The IDE isn’t dead. Cloud agents are interesting. Autonomous task completion is real. But “the IDE is over” is a thesis, not a fact. Most developers still spend most of their time in an editor. Building the best experience there still matters.

Open source is a hedge. Open source tools give you the option to fork, migrate, or run your own instance. Open weight models give even more flexibility in terms of what inference provider you use as the models improve every week.

The AI coding landscape is consolidating: closed ecosystems are getting bigger and more vertically integrated. The alternative is open, model-agnostic tooling that doesn’t tie your workflow to any single provider — and that’s what we’re building.

---

Kilo is the most popular open source AI coding agent. Install the VS Code extension or try Kilo Cloud.