Cursor vs Cognition: Opposite Takes on Agent Search

Source: https://x.com (tweet) Author: unknown Date: 2025-2026

Summary

Two leading coding agent companies have opposite architectures for code search: Cursor uses custom embeddings trained on agent traces (12.5% accuracy improvement); Cognition rejects embeddings entirely, instead training models to use grep with 8x parallel tool calls. Both work.

Key Claims

  • Cursor: Custom embeddings trained on agent traces → +12.5% accuracy
  • Cognition: No embeddings; models trained to use grep with 8x parallel tool calls
  • Both approaches can succeed — architecture is not settled
  • The divergence is philosophically interesting: symbolic (grep) vs. neural (embeddings)

Concepts

  • Coding Agents — search strategy as core architectural decision
  • Agent Memory — embeddings as a form of indexed agent memory