Beyond grep: The case for a context-rich AI coding harness
Ars Technica examined how AI coding assistants differ not just in their underlying models but in the "harness" software that governs what those models can see and do within a codebase. Anthropic's Claude Code deliberately keeps this harness lean, avoiding pre-built structured context systems, on the reasoning that models are improving so rapidly that heavy, opinionated tooling risks becoming a constraint rather than a help. Rival Augment Code has taken the opposite approach, betting that richer context retrieval gives its agents an edge, which highlights a genuine industry-wide split in how AI coding tools are being engineered.
The piece draws on interviews with Cat Wu, Anthropic's head of product for Claude Code, and Vinay Perneti, Augment Code's VP of Engineering. Wu said Claude Code favours simple, grep-based context retrieval, noting that evaluations show no measurable benefit from more elaborate approaches, and prefers letting developers bolt on their own tools if needed. Perneti described Augment Code's alternative: pre-indexing an entire repository using an embedding and retrieval model plus a vector database, enabling sub-millisecond semantic retrieval of conceptually relevant code, which the company argues better serves agents working within limited context windows. Other harnesses mentioned include OpenAI's Codex, Google's Antigravity, OpenCode and Cursor.
- AI coding tools differ in "harness" design, not just the underlying model
- Claude Code favours a lean, grep-based approach over pre-built context
- Augment Code instead pre-indexes code with embeddings for semantic retrieval
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Advocates of a lean, grep-based harness argue that AI coding models are advancing so quickly that today's clever retrieval systems risk becoming tomorrow's dead weight, embedding assumptions about model limitations that newer, more capable models simply outgrow. They value simplicity, transparency and flexibility, preferring to let developers attach their own tools as needed rather than lock users into a heavyweight, opinionated system whose benefits, according to their own evaluations, have not been clearly demonstrated over straightforward search.
The case against
Proponents of a context-rich harness contend that pre-indexing a codebase with embeddings and a vector database allows an agent to retrieve conceptually relevant code almost instantly, which matters greatly when a model's context window is limited and every token counts. They value engineering for the constraints models actually face today, arguing that investing in sophisticated retrieval infrastructure gives agents a meaningful, measurable edge in navigating large, complex repositories rather than waiting for future model improvements to close the gap.