Cursor engineering blog

Cursor explains how its agent harness reduced token costs by 7% without lowering quality. The post details system-prompt trimming, on-demand tool loading, explicit cache breakpoints, compact file reads, and more selective subagent use, with measured effects on token usage and cache misses.

SpaceXAI introduces Grok 4.7, a model optimized for extended coding and knowledge-work tasks through a larger base model and longer reinforcement-learning run. The post reports benchmark results, safety and cybersecurity evaluations, API pricing, and availability across Grok, Cursor, and third-party platforms.

Vicent Martí explains why hosting Git repositories at scale is difficult, covering packfile layout, distributed filesystems, replication, and consistency challenges. He presents Cursor’s Continuity architecture, which uses an S3-backed write-ahead log, linearizable pushes, elastic replication, and WAL-driven compaction to scale Git storage reliably.

The post announces Grok 4.6. The body is unavailable, so no technical details or implementation takeaways can be verified from the title and URL alone.

The post appears to explain how Cursor Router selects an appropriate AI model for each task. The body was unavailable, so the technical criteria, routing architecture, and trade-offs could not be verified.

The post introduces Cursor’s open-source mixture-of-experts megakernel, designed for NVL72 GPU systems. The body is unavailable, so implementation details and measured benefits cannot be verified.

The post outlines how Cursor set up its cloud agent environment. Because the body is unavailable, the implementation details and any transferable lessons cannot be verified.

The post examines agent swarms and the changing economics of AI models. The body was unavailable, so this summary is based on the title alone.

Cursor announces the Grok 4.5 model card, which covers the model’s capability benchmarks and safety evaluations. The page provides a link to the full documentation but contains no further technical detail.

The post announces Grok 4.5, but its body was unavailable, so no technical details or implementation takeaways could be verified from the title alone.

The post examines how reward hacking can overwhelm measured model-intelligence gains. The body was unavailable, so this summary is based on the title alone.

The post discusses governing the autonomy of AI agents through an auto-review mechanism, but its body was unavailable for verification.

The post discusses lessons from building cloud agents, but its body is unavailable, so specific techniques and findings cannot be verified.

The post appears to announce Composer 2.5 from Cursor, but its body was unavailable, so no specific features, implementation details, or technical takeaways could be verified from the title and URL alone.