Cursor AI Editor Goes Agent-First After $60B SpaceX Takeover

From autocomplete to an agent control room

The AI code editor has turned into something closer to a dispatch desk. Coverage of Cursor, the editor built by Anysphere, describes a product that changed its basic design over the past year. Its main job is no longer predicting the next line of code. Its main job is now coordinating software agents that plan, write, test and submit changes on a developer's behalf. Rival Windsurf took the same path and went further, giving up its name to become the desktop front end of Cognition's Devin agent platform. Meanwhile, prompt-to-app builders such as Lovable and Bolt.new are going after a different customer, people who never open an IDE at all, and investors are paying heavily for that market too.

The overall picture is clear. The race for AI coding tools is no longer about which editor has the smartest autocomplete. It is about who owns the layer where autonomous agents get assigned work and where their output gets reviewed. That shift explains the product redesigns, the large valuations and the consolidation reported across the coverage.

What Cursor actually shipped

The release history is consistent across reviews, even where individual dates differ slightly. Cursor 2.0 arrived in October 2025. It let developers run several AI agents in parallel2, and it introduced Composer, the company's own coding model, which one reviewer found to be roughly twice as fast as Claude Sonnet 4.55. The same reviewer said that release made Cursor feel less like VS Code with AI added and more like an agent workbench that also happens to be an editor5.

Version 2.4, released in January 2026, added subagents. These are independent agents with their own context windows that a primary agent can launch to handle separate subtasks in parallel3. It also added Skills, which are task-specific workflows the agent loads only when needed. Rules, by contrast, set standing coding standards2. One tutorial says that by that month Cursor had about 10 million active developers, up from around 4 million the previous June3.

The biggest structural change came with Cursor 3.0 on April 2, 2026. It added an Agents Window for running many agents across repositories and environments, whether local, in git worktrees, in the cloud or over remote SSH7. It also brought commands such as /worktree, which isolates an agent's edits in a separate checkout, and /best-of-n, which runs the same task on several models at once7. Releases 3.3 and 3.5 followed in May. The first added "Build in Parallel" subagents, Composer 2.5 and native Jira integration. The second added Cloud Agents, which run in isolated virtual machines with their own terminal, browser and desktop access and are billed separately from the seat price7. Bugbot, Cursor's pull-request reviewer, can now start its own cloud agent to test and propose a fix. One guide puts its resolution rate at about 80%7.

Some sources disagree on details. One tutorial dates Cloud Agents to version 3.5 and says they require the Business plan3, while another guide simply calls them new in 2026 and describes starting them from a browser, a phone or Slack4. A tutorial says Agent Skills were nightly-only in early 20263, while another guide treats Skills as an official 2.4 feature2. These are timing differences, not conflicting accounts. The direction of the product is the same in all of them.

The coverage also lists the weaknesses. Long agent sessions lose reasoning quality, large monorepos strain the indexer, and multi-file refactors can produce messy pull requests that still need a human to review them7. Older reviews mention similar problems, including context "forgetfulness" and generated code being placed in the wrong spot1.

The money: a $60 billion exit

Cursor's product changes came alongside an unusual financial run. Anysphere's valuation went from $400 million at its August 2024 Series A to $9.9 billion in June 2025 and $29.3 billion at a $2.3 billion Series D in November 20251114. One investor-focused account dates that Series D to January 202616, but most reporting places it in November.

Revenue rose even faster. Reported annualized revenue grew from $100 million in January 2025 to $1 billion by November 2025, $2 billion by February 2026 and about $4 billion by June 202612. In April 2026, an oversubscribed round of at least $2 billion at about a $50 billion valuation was reported, co-led by Andreessen Horowitz and Thrive Capital with Nvidia as a strategic investor1520.

The accounts of what happened to that round differ. One valuation guide says the Series E talks were dropped after SpaceX took an option to buy the company for $60 billion, with a walk-away payment of about $10 billion17. Dealroom, by contrast, reported in July that the $50 billion round had been raised18. Reporting agrees on the larger outcome. SpaceX announced an all-stock deal valuing Anysphere at $60 billion on June 16, 2026, four days after its own IPO12. One tracker reports that the deal closed on August 14, 202611. Against roughly $4 billion in annualized revenue, that price works out to about 15 times revenue12.

The main consequence is that Cursor, which had been the leading independent AI-native editor, is no longer independent. Comparison tables written as recently as August still listed Anysphere as the owner, separate from any larger company31. That is no longer accurate.

Windsurf becomes Devin Desktop

Windsurf went through even more turmoil. Its planned roughly $3 billion sale to OpenAI fell apart on July 11, 2025. Google then reportedly paid about $2.4 billion to license the technology and bring CEO Varun Mohan and other staff into DeepMind. Cognition, the maker of the Devin agent, bought the remaining company, product and brand on July 14, 20253538. One review instead dates the acquisition to December 2025 at about $250 million31, but most of the reporting supports the July 2025 timeline.

Under Cognition, the editor received the company's own SWE-series models and a codebase navigation feature called Codemaps3138. On March 19, 2026, Cognition dropped Windsurf's credit-based pricing and raised the Pro plan from $15 to $20 a month, the same price as Cursor3139. On June 2, 2026, an over-the-air update renamed the product Devin Desktop. Plans, extensions and settings carried over, and windsurf.com now redirects to Devin's site3435.

The most telling change is to the default view. The app now opens on an Agent Command Center, a Kanban board of local and cloud agent sessions, and the code editor has become a tab inside it34. Cascade, the agent that made Windsurf's reputation, has been replaced by Devin Local, a Rust rewrite that Cognition says is up to 30% more token-efficient3440. Sources give different dates for Cascade's retirement, from support pledges through July 202634 to a changelog entry recording its removal on September 8, 202637. Devin Desktop also supports the Agent Client Protocol, so third-party agents such as Claude Code and Codex can run inside it3240.

Security researchers have raised concerns. OX Security reported a prompt-injection flaw that allowed remote command execution in an April 2026 Windsurf build, and Cognition's changelog lists a Restricted Mode bypass that was fixed on September 16, 202637.

Lovable and Bolt: the no-IDE flank

At the other end of the market, prompt-to-app builders are competing for people who would never open Cursor or Devin Desktop. Lovable's growth has been the fastest. It raised a $400 million Series C on August 12, 2026 at a $13.3 billion valuation, led by Menlo Ventures and co-led by EQT's Scaleup Europe Fund2328. That valuation roughly doubles the $6.6 billion from its December 2025 Series B22. The company says users have built more than 60 million projects and that apps built on Lovable receive more than 900 million visits a month2326.

Figures for Lovable's recurring revenue vary by source and date. Reported numbers include about $400 million in February 202621, $500 million by June24 and an estimate near $600 million by the time of the Series C25. One analysis cautions that ARR is not audited recognized revenue23. Bolt.new, built by StackBlitz, is much smaller, with its most widely cited figure being $40 million in ARR within about six months of launch2130. One comparison says Bolt has narrowed the backend-feature gap through Bolt Cloud's built-in databases, authentication and hosting30.

The reading

Taken together, the coverage points to the IDE becoming a supervisor's console. Cursor's Agents Window and Devin Desktop's Command Center were designed separately but solve the same problem: giving developers one place to dispatch, monitor and merge the work of many agents. Both companies have also put money into their own models, Composer at Cursor and SWE-1.x at Cognition, to reduce their reliance on frontier labs531.

Price is no longer a point of difference, since both editors start at $2039. Competition now depends on how well the agents work, how the tools fit into enterprise workflows and whether developers trust them. That last factor matters more as agents get terminal and browser access. Lovable's valuation shows that a large part of the market is not developers at all. The open question is whether apps built from prompts and pull requests generated by agents can hold up in production at the scale investors are paying for, and the coverage does not yet answer it.