AI 工具链（harness）对性能的影响比模型本身更大。
Endor Labs 在同一周内用两套工具链运行了相同的模型。OpenAI 的 GPT-5.5 在其原生 Codex 工具链中功能正确率为 61.5%，而在 Cursor 工具链中为 87.2%，仅因运行时不同就产生了 25.7 个百分点的差距。Anthropic 的 Opus 4.7 在 Claude Code 中得分 87.2%，在 Cursor 中为 91.1%。
两款前沿模型在竞争对手的工具链中表现都优于自家发布的工具链。
工具链是 AI 技术栈中的关键支点。它们影响成本、质量和准确性。
在 OpenRouter 上，输入 token 占大语言模型流量的 86-98%。输出每 token 的成本是输入的 5 倍，但输入仍然占据账单大头，因为输入量要大得多。一年前，我在《Hungry, Hungry AI Model》中曾思考过这个比例问题。从业者当时说 95% 输入 / 5% 输出。在规模化场景下，他们是对的。
因此，控制输入成本是一个极具价值的命题。模型本身无法控制这些成本，但工具链可以。工具链决定发送哪些上下文。这些上下文中有很大一部分在每次查询之间是重复的。智能地缓存这些内容可以节省 40-80%。一项针对 500 个长时程智能体会话的研究发现，成本降低 41-80%，首 token 延迟加快 13-31%，且节省效果从 500 token 到 50,000 token 的提示词呈线性增长。胜出的方案是只缓存稳定的前缀，并将动态内容放在缓存断点之后。
工具链还负责信息检索：要读取的代码、风格指南、投资简报中需要分析的部分。内容越简洁精准，成本就越低。
Cursor 的工具链在技术上与 Claude Code 逐项对标：动态工具获取、基于优先级的提示词前缀组装、双层缓存。这就是为什么 Opus 4.7 在 Cursor 中的得分高于 Claude Code，也是为什么 GPT-5.5 在 Codex 之外的功能正确率几乎翻倍。
捆绑方案并非没有可取之处。Claude Code 将缓存命中率视为可用性指标，在真实会话中命中率约为 96%，在同一版本的用户之间共享系统提示词缓存，并以 99% 的字节一致性构建分叉子智能体，节省 90% 的成本。
协同设计 harness、缓存 API 与模型，确实能换来真正的缓存纪律。但这份纪律存在于 harness 之中，而第三方 harness 同样可以达到同等水平，Cursor 的数据便证明了这一点。
Harness 已摆脱了“模型包装器”的批评之声；它如今是一位骑师，将 AI 的性能推向比培育者所想象的更远的地方。
-
Endor Labs，“GPT-5.5 在 Agent 安全联赛中凭借 Cursor（而非 Codex）创下代码安全新纪录”，2026 年 4 月 27 日，https://www.endorlabs.com/learn/gpt-5-5-sets-a-new-code-security-record-with-cursor-not-codex-in-agent-security-league。Agent 安全联赛基准测试，基于卡内基梅隆大学的 Open SusVibes 框架。↩︎
-
Tomasz Tunguz，“饥饿的 AI 模型”，2025 年 7 月 8 日，https://tomtunguz.com/input-output-ratio/ ↩︎
-
Tomasz Tunguz，“Harness 是新的战场”，https://tomtunguz.com/the-harness-is-the-new-battleground/ ↩︎
-
Lumer 等人，“不要破坏缓存：面向长周期智能体任务的提示词缓存评估”，arXiv:2601.06007，2026 年。https://arxiv.org/abs/2601.06007 ↩︎
-
Anthropic，“构建 Claude Code 的经验教训：提示词缓存就是一切”，2026 年 4 月，https://claude.com/blog/lessons-from-building-claude-code-prompt-caching-is-everything；交换税数据来自 Digital Applied，“提示词缓存经济学：缓存优先的智能体设计”，2026 年 7 月，https://www.digitalapplied.com/blog/prompt-caching-economics-cache-first-agent-architecture-2026。 ↩︎
AI harnesses have more impact on performance than the models.
Endor Labs ran the same models through two harnesses in the same week. OpenAI’s GPT-5.5 scored 61.5% functional correctness in its native Codex harness & 87.2% in Cursor’s, a 25.7-point swing from the runtime alone. Anthropic’s Opus 4.7 scored 87.2% in Claude Code & 91.1% in Cursor.1
Both frontier models performed better in a competitor’s harness than in the one their maker ships.
Harnesses are a fulcrum in the AI stack. They influence cost, quality, & accuracy.
Input tokens are 86-98% of LLM traffic on OpenRouter. Output costs 5x input per token, but input still dominates the bill because there is so much more of it. One year ago I wondered about this ratio in the Hungry, Hungry AI Model.2 Practitioners said 95% input/5% output. At scale, they were right.
Controlling input costs is therefore a valuable proposition. The model doesn’t control those costs, but the harness does.3 Harnesses decide which context to send. Much of that context repeats from query to query. Caching it intelligently can save 40-80%. A study across 500 long-horizon agent sessions found 41-80% cost reduction & 13-31% faster time-to-first-token, with savings scaling linearly from 500- to 50,000-token prompts.4 The winner was caching only the stable prefix & placing dynamic content after the cache breakpoint.
Harnesses also perform information retrieval: the code to read, the style guide, the section of an investment brief to analyze. The more concise & precise, the lower the cost.
Cursor’s harness matches Claude Code technique-for-technique: dynamic tool fetching, priority-based prefix assembly, two-tier caching. That is why Opus 4.7 scored higher in Cursor than in Claude Code, & why GPT-5.5 nearly doubled its functional-correctness score outside Codex.
The bundled case is not without merit. Claude Code treats cache hit rate as an uptime metric, hits about 96% in real sessions, shares the system-prompt cache across users on the same version, & builds forked sub-agents at 99% byte-identity for 90% savings.5
Co-designing the harness, cache API, & model buys real cache discipline. But the discipline lives in the harness, & a third-party harness can match it, as Cursor’s numbers show.
The harness has shed its critique as a model wrapper ; it’s a jockey pushing AI performance further than the breeder imagined.
-
Endor Labs, “GPT-5.5 Sets a New Code Security Record with Cursor, not Codex, in Agent Security League,” Apr 27, 2026, https://www.endorlabs.com/learn/gpt-5-5-sets-a-new-code-security-record-with-cursor-not-codex-in-agent-security-league. Agent Security League benchmark, based on Carnegie Mellon’s Open SusVibes framework. ↩︎
-
Tomasz Tunguz, “The Hungry, Hungry AI Model,” July 8, 2025, https://tomtunguz.com/input-output-ratio/ ↩︎
-
Tomasz Tunguz, “The Harness Is the New Battleground,” https://tomtunguz.com/the-harness-is-the-new-battleground/ ↩︎
-
Lumer et al., “Don’t Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks,” arXiv:2601.06007, 2026. https://arxiv.org/abs/2601.06007 ↩︎
-
Anthropic, “Lessons from building Claude Code: Prompt caching is everything,” Apr 2026, https://claude.com/blog/lessons-from-building-claude-code-prompt-caching-is-everything; swap-tax figures from Digital Applied, “Prompt Caching Economics: Cache-First Agent Design,” Jul 2026, https://www.digitalapplied.com/blog/prompt-caching-economics-cache-first-agent-architecture-2026. ↩︎