cursor
Cursor — AI-first IDE (the company); developer of the Composer coding-model series. Already referenced in two sources via the file-backed MCP descriptions A/B test (2026-04-27-agent-principles-architecture-engineering) and now via the Composer 2 technical report (2026-04-27-llm-training-principles-paths-practices).
What this wiki currently knows
Products / models referenced
- Composer 2 — Cursor’s in-house coding model. Public engineering reference for two patterns: real-time RL feeding production traffic back into training, and self-summarization as a graded subtask.
Techniques attributed
- Real-time RL — production traffic continuously updates training signals rather than waiting for the next offline retrain pass. Cited as one example of the offline / online training boundary closing (post-training, 2026-04-27-llm-training-principles-paths-practices §“前沿模型发布后”).
- Summary-quality reward — long coding sessions have model-generated summaries injected back into context; if the summary distorts what happened, the rest of the session compounds the drift. Cursor incorporates summary fidelity into the reward, so summary becomes a graded subtask, not a free output.
- GRPO-style RL infrastructure — the article notes Cursor’s RL stack converged on a GRPO-class algorithm (no separate value network), the same direction as DeepSeek (grpo).
- MCP file-backed descriptions A/B test — older reference from 2026-04-27-agent-principles-architecture-engineering: file-backed tool descriptions cut total token usage 46.9% on MCP-using tasks. Tracked on model-context-protocol and claude-skills.
Stance
Cursor is the wiki’s reference for “coding agent in production”: their public engineering posts are short on theoretical claims and long on what tactics moved which numbers in real workflows. The Composer 2 work is a useful counterweight to lab-style training reports (DeepSeek-V3, Llama 3) — same modern post-training stack, applied to an active product with continuously-arriving user traffic.