Researchers built an AI agent that searches text the way a sysadmin would: with grep, not a search index.
The system, called GrepSeek, skips the usual retrieval setup where a model queries a prebuilt, ranked index. Instead it treats the raw corpus as its environment and finds evidence by running shell commands directly against the text. The team trained it in two stages: first imitating verified search trajectories from an answer-aware "Tutor" and an answer-blind "Planner," then refining the policy with reinforcement learning (Group Relative Policy Optimization). To make shell-based search fast enough to be useful, they built two execution tricks: one cuts search latency up to 77 times on a 14GB, 21-million-document corpus, and another parallelizes search across shards for up to a 7.6 times speedup, both without changing the results you'd get from sequential search.
Across eight open-domain QA benchmarks, GrepSeek beat the best baseline by a statistically significant 5.7% relative margin. That's a meaningful jump for an approach that skips the index-building step most retrieval systems treat as mandatory. It suggests letting an agent poke around raw text with command-line tools can beat handing it whatever a ranker decided it should see.
Indexing has been retrieval's default assumption since before LLMs existed. GrepSeek is a reminder that defaults are worth re-testing once the thing doing the searching can write its own queries.