Cursor and Opus prototypes now auto‑generate 1,000+ lines of code

Writing code with artificial intelligence now means asking a model to plan work across a repository, not just finish a function. Cursor and Anthropic both now describe tools built for longer, multi-file coding jobs. (cursor.com) (anthropic.com) Cursor introduced Plan Mode on October 7, 2025. The company said its agent can research a codebase, review docs, ask clarifying questions, and turn that work into an editable Markdown plan with file paths and code references. (cursor.com) Cursor said developers can then build directly from that plan, and that the editor will suggest Plan Mode automatically for complex tasks. The company also said its agents can now run “significantly longer” than before. (cursor.com) Anthropic made a parallel claim on April 16, 2026, when it launched Claude Opus 4.7. The company said users can hand off harder coding work that previously needed close supervision, because the model handles long-running tasks with more consistency and checks its own output before replying. (anthropic.com) That is a different job from code completion. The model first maps the work, then edits files, then verifies what it changed, which is closer to a junior engineer following a ticket than to autocomplete in an editor. (cursor.com) (anthropic.com) Cursor has also been testing what happens when that pattern scales out. In a January 14, 2026 research post, the company said hundreds of concurrent agents wrote more than 1 million lines of code across 1,000 files on a single project. (cursor.com) The company said its first coordination scheme broke down because agents held locks too long, forgot to release them, or avoided hard tasks. Cursor then split the system into planners that create tasks and workers that execute them. (cursor.com) Anthropic has reported a similar structure. In a February 5, 2026 engineering post, researcher Nicholas Carlini said 16 Claude agents working in parallel produced a 100,000-line Rust-based C compiler over nearly 2,000 Claude Code sessions at a cost of $20,000. (anthropic.com) Those experiments also show the tradeoffs. Anthropic said on April 23 that changes meant to reduce latency and verbosity had hurt Claude Code quality, and the company reverted them after users reported worse coding results. (anthropic.com) The practical change for software teams is that review shifts upward. Instead of checking whether the assistant wrote a loop correctly, engineers are increasingly checking whether the plan, tests, and verification steps match the original brief. (cursor.com) (anthropic.com) The headline number in social posts is 1,000 lines of code, but the more durable change is that these tools are being built to own an entire coding task from plan to test. That makes the bottleneck less typing and more supervision. (cursor.com) (anthropic.com)