Cursor co-founder Swallie discusses the release of Composer 1.5, a model trained heavily through reinforcement learning positioned between Sonnet 4.5 and Opus 4.5 in capability. Key topics include why Cursor builds its own models (to bake tool-use and semantic search capabilities directly into training), the infrastructure challenges of running millions of sandboxes for RL, and the current limitations of cloud agents versus local agents. He explains that cloud agent adoption needs to grow 1000x, which requires models that can test their own code rather than just UI tweaks. The conversation covers context window management via RL-trained self-summarization, the browser-building multi-agent experiment that produced thousands of commits autonomously, and a vision of 'self-driving codebases' where AI autonomously handles security, tech debt, and bug-fixing. On the future of engineering, he notes that manual coding has largely been replaced among top engineers and expects 1-2 capability jumps every six months, with developers increasingly taking on managerial roles directing agents.
Sort: