INTRODUCING SWE-GREP AND SWE-GREP-MINI: RL FOR MULTI-TURN, FAST CONTEXT RETRIEVAL

Summary from the TLDR AI newsletter

Cognition trained specialized models that retrieve code context 10x faster than frontier models by using RL to issue up to 8 parallel tool calls per turn instead of the sequential searches typical agents use. Speed matters more than the industry realizes because it allows developers to stay in a flow state.