SWE-Grep and SWE-Grep-Mini: RL for Fast Multi-Turn Context Retrieval (cognition.ai)

Cognition/Windsurf announced SWE-grep and SWE-grep-mini, lightweight agentic models trained with multi-turn reinforcement learning to do highly parallel codebase context retrieval. Instead of the usual slow, serial tool calls, SWE-grep natively issues up to 8 parallel tool calls per turn across a maximum of 4 turns (3 exploration + 1 answer), using a restricted safe toolset (grep, read, glob, etc.). Training used a Cognition CodeSearch Eval of real repos and queries, with a weighted F1 reward that prioritizes precision on file- and line-level retrieval; SWE-grep was distilled to SWE-grep-mini and further tuned. To get stable, unbiased RL updates when rollouts include many environment tokens, the team applied per-sequence importance sampling, a leave-one-out baseline to reduce variance, and other engineering tweaks. Results show SWE-grep models match or outperform frontier coding models for retrieval while reducing end-to-end latency by about an order of magnitude. SWE-grep powers Windsurf’s new Fast Context subagent (auto-triggered in normal use or forced with Cmd+Enter) and is available in a playground/demo; benchmark comparisons (internal runners vs. Claude Code and Cursor CLI) demonstrate much lower latency and faster completion of difficult coding tasks when used as a subagent in Windsurf’s Cascade (Sonnet 4.5 main model). The work addresses the speed–intelligence tradeoff in coding agents, prioritizing rapid context delivery to keep developers “in flow” and paving the way for wider deployment across Windsurf products.