Is grep all you need for agentic search?

Mintlify's approach of replacing RAG with grep over a virtual filesystem sparked a broader discussion about agentic search design. The post argues that 'dumb retrieval tools' like grep can work for agents when paired with the right constraints. Key concepts include structured tool signatures that limit agent choices, and a harness pattern with inner and outer loops: the inner loop runs the LLM until it's satisfied with its tool calls, while the outer loop validates results against programmatic quality criteria (hooks). Hooks act as automated feedback, telling the agent to try harder when results don't meet quality bars. However, the author cautions that naive retrieval has real limits — token costs accumulate with many tool calls, and complex ranking factors (recency, popularity, text relevance) become hard to model without proper search infrastructure. The conclusion is that while grep-based agentic search is instructive for understanding harness design, production systems will eventually need proper retrieval tooling.