grep を超えて：コンテキスト重視の AI コーディングハーネスの必要性Beyond grep: The case for a context-rich AI coding harness

For decades, grep has been the workhorse of code navigation, letting developers locate strings, symbols, and patterns across sprawling repositories with a single command. As AI coding assistants become embedded in everyday workflows, a growing argument holds that plain text search is no longer sufficient to feed these models the information they need. The case being made is that the surrounding scaffolding, often called the "harness," matters as much as the underlying model, and that harnesses built around simple keyword matching leave significant accuracy and productivity on the table.

The core problem is that grep and its faster successors, such as ripgrep, operate on literal text. They can tell you where a token appears, but not what it means. A search for a function name returns every occurrence, including comments, unrelated strings, and definitions in files that are never actually reached at runtime. For a human, filtering that noise is often trivial. For a large language model working within a limited context window, indiscriminately dumping search results consumes valuable tokens and can bury the relevant signal. The article argues that this mismatch degrades both the quality of generated code and the efficiency of the interaction.

A context-rich harness, by contrast, aims to give the model a structural understanding of a codebase rather than a flat list of matches. This typically draws on techniques that have existed in developer tooling for years but are now being repurposed for AI. Language servers built on the Language Server Protocol can resolve where a symbol is defined and every place it is genuinely used. Abstract syntax trees and tools like tree-sitter allow code to be parsed into meaningful units, so a model can be handed a complete function or class rather than an arbitrary window of lines. Call graphs and dependency analysis can trace how components relate, helping the assistant follow the actual flow of logic instead of guessing from names alone.

Retrieval is another central piece. Many current systems combine lexical search with semantic search, using vector embeddings to find code that is conceptually related even when it shares no keywords. The likely direction described here is a hybrid approach: fast text search to anchor exact matches, semantic retrieval to surface related context, and structural analysis to assemble a coherent picture before anything reaches the model. The goal is to present the assistant with the smallest set of highly relevant material, a practice that has become known more broadly as context engineering.

This debate is playing out across a crowded field of developer tools. Editors and assistants such as GitHub Copilot, Cursor, and various open-source agents have invested heavily in indexing and retrieval layers, and the emergence of standards like the Model Context Protocol reflects an industry effort to standardize how models connect to external sources of information. Agentic coding tools that can run commands, read files, and iterate over multiple steps make the quality of the harness even more consequential, because errors introduced early can compound across a long chain of automated actions.

It is worth noting some caveats. Richer context comes at a cost. Building and maintaining code indexes, running language servers, and computing embeddings all add latency, infrastructure, and complexity, and indexes can drift out of date in fast-moving repositories. Semantic retrieval is not infallible and can return plausible but irrelevant results. Grep itself remains fast, dependable, and language-agnostic, which is why it is likely to stay a component of these systems rather than disappear entirely. The stronger claim in the piece is not that text search is useless, but that relying on it alone appears to be a bottleneck for AI-assisted development.

The broader takeaway is that improvements in AI coding may increasingly come from the tooling around the model rather than from raw model capability. As context windows grow and models improve, the question of what to place inside that window, and how to structure it, becomes a distinct engineering discipline. For teams evaluating AI coding assistants, the argument suggests looking closely at how a tool gathers and organizes context, since that layer may prove to be a more reliable differentiator than benchmark scores. Whether the industry converges on shared standards or fragments into competing approaches remains to be seen, but the shift from searching text to understanding code appears to be well underway.