ByteAsk 定位引擎砍掉 78% token：C/C++ 修复成本降 68.1%
写 C++ 的智能体烧钱主要烧在读代码上，ByteAsk 先把该改哪儿标好再让模型动手，token 直接从 1795K 降到 395K，成本省了 68.1%。天天跑 agent 修 bug 的团队可以试试，账单应该能明显瘦下来。
ByteAsk 推出一项针对 C/C++ 编码智能体的上下文定位引擎，在同一组测试下修复 bug 数量不变，token 消耗减少 78%，成本降低 68.1%。
其机制是在大模型开始读代码之前，先按构建系统的视角扫描代码库，标出修复应该落在哪些位置，还会标记需要同样改动的兄弟位置（sibling sites）；智能体随后只需针对覆盖这些位置的测试验证补丁，不必通读全库。这使得每次运行的池化 token 从 1,795K 降到 395K。
文中指出，智能体修 C++ 的大部分开销本来花在“找代码”而非“改代码”上，这一前置定位在冷门项目和高难度问题上也能实现约 80% 的节省，对被 token 账单压着的 C/C++ 开发团队是直接的成本优化。
Another great launch to reduce token cost.
Most of what you pay an AI agent to fix C++ goes to finding the code, before any fixing. @ByteAsk removed 78% of the tokens and fixed exactly as many bugs and at 68.1% lower cost.
ByteAsk is a C and C++ coding agent that finds where a fix belongs before the AI model starts reading. Before the model starts, ByteAsk’s context engine reads the code as the build sees it and marks where the change belongs.
It also flags sibling sites needing the same edit, and the agent checks its patch against only the tests covering them.
That upfront map cut pooled tokens from 1,795K to 395K per run, with about 80% savings on little-known projects and hard issues.
This report is categorized under AITOP's "Products" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include ByteAsk, C++, coding agent, token成本.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
