ByteAsk: Build-Aware Context and a Checking Harness for C++ Coding Agents

Key findings

1. 01Same fixes as Codex CLI on the same model: 129 of 144 runs fixed by each.
2. 0263–80% fewer tokens and 59–70% lower cost in every task set.
3. 03The biggest savings come on little-known and hard C++ projects: about 80% fewer tokens.

Abstract

Large language model (LLM) coding agents have changed how people write software: a developer describes a change in plain language and the agent finds, edits, and tests the code. Most of this progress, however, has concentrated on a few widely used languages, Python above all. C and C++, which still underpin operating systems, browsers, databases, and embedded software, remain far less explored, even though their large, build-dependent codebases are exactly the setting in which agents struggle to find where a change belongs. We propose ByteAsk, a coding agent built for C and C++. At its core is a context engine that understands a codebase the way the project's own build does. Before the model starts, the engine tells it where the change belongs, which related places may need the same change, and which tests cover them; ByteAsk then checks the change against only those tests. We compare ByteAsk with OpenAI's Codex CLI running the same underlying model, on recent issues from popular and little-known C++ projects and on a small set of hard issues. Both agents fix the same number of runs in every set (129 of 144 each, across the two main sets), but ByteAsk uses 63–80% fewer tokens and costs 59–70% less.

Cite this paper

Pratyush Saini and Anirudha Kulkarni. ByteAsk: Build-Aware Context and a Checking Harness for C++ Coding Agents. ByteAsk Research, preprint, November 2026. byteask.ai/research/byteask-system-cpp

@misc{saini2026byteasksystem,
  title        = {ByteAsk: Build-Aware Context and a Checking Harness for C++ Coding Agents},
  author       = {Saini, Pratyush and Kulkarni, Anirudha},
  year         = {2026},
  month        = nov,
  institution  = {ByteAsk Research},
  howpublished = {\url{https://byteask.ai/research/byteask-system-cpp}},
  note         = {Preprint. Code and data: \url{https://github.com/ByteAsk/byteask-system-cpp}}
}