Building against the big labs that are trying to eat you如何在试图吞噬你的大型实验室面前构建产品
The ongoing doomer sentiment on big model companies eating your product / startup is getting a little exhausting. Startups and builders still have natural structural advantages, and these can be used to carve out a niche and win.关于大型模型公司会吞噬掉你的产品或初创企业的悲观论调，听得实在让人有些疲惫。初创公司和开发者依然拥有天然的结构性优势，只要善加利用，完全可以开辟出利基市场并取得胜利。
So I'm sharing lessons I learned building the Shortcut agent when Claude for Excel is our competitor. 因此，我想分享一下在 Claude for Excel 作为竞争对手的情况下，我开发 Shortcut 智能体所学到的经验。
Your harness is better because your life depends on it你的架构更好，因为你的身家性命系于此
For the labs, the main quests are training the smartest model, recursive self-improvement, and having enough enterprise revenue to do the first two. Everything else is secondary.对于大型实验室而言，他们的主要任务是训练最聪明的模型、实现递归自我改进，并获得足够的企业营收来支撑前两项工作。除此之外的一切都是次要的。
For me, agentic performance on spreadsheet tasks is the whole of Shortcut existence. So I'm going to beat competitors by obsessing over the harness. Prompts, tools, and context are ruthlessly refined until they are beautifully minimalistic and general. For our internal evals, which reflects the broad swaths of finance work done by our users (to be publicly released), we are 40% cheaper and 17% more accurate than Claude for Excel even when we standardize on the same base model (Opus 4.8).对我来说，电子表格任务中的智能体性能就是 Shortcut 生存的全部。因此，我将通过对架构的极致打磨来击败竞争对手。提示词、工具和上下文经过了无情的精简，直到它们变得极其简洁且通用。根据我们内部的评估（反映了用户进行的大量金融工作，即将公开发布），即便我们使用相同的基准模型（Opus 4.8），我们的成本也比 Claude for Excel 低 40%，准确率高 17%。
We make roughly half the tool calls (37 vs 61 per task) because we have more efficient tools and are somehow more code-execution-pilled than Claude for Excel.我们调用的工具次数大约只有他们的一半（每个任务 37 次对 61 次），因为我们的工具效率更高，而且在代码执行方面似乎比 Claude for Excel 更“激进”。
We take about the same number of turns they do - 36 vs 34 - but each of our turns uses a much lighter context: 3.7M input tokens per task against their 7.1M. After we break the per-task cost down by token class, our output token cost matches theirs, but they pay far more to re-read a heavier context on every turn. 我们的交互轮数与他们大致相当（36 对 34），但我们每一轮使用的上下文要轻量得多：每个任务的输入 Token 为 370 万，而他们是 710 万。在按 Token 类型拆解单任务成本后，我们的输出 Token 成本与他们相当，但他们为了在每一轮中重新读取更沉重的上下文，付出了高昂的代价。
A leaner context is cheaper to run and smarter to reason over, so we win on both cost and accuracy. 更精简的上下文运行成本更低，推理也更聪明，所以我们在成本和准确率上都赢了。
Your harness is better by definition你的架构在定义上就更胜一筹
Why? Because your domain is smaller than theirs. Claude Code and Codex ships 30+ tools each because they have to. Their tail is larger than your entire customer base, so they must support features for each of these use cases. But their capitulation to breadth floods their agent's context with additional instructions, tools, and skills, whereas you can trim all this fat to focus on 2-3 killer use cases. As an example, Pi's stripped-down harness beats Claude Code and Codex on , and it's not because the folks at Anthropic and OpenAI don't know what they're doing. CC and Codex serve many millions of customers and Pi is only serving devs and evals.为什么？因为你的领域比他们小。Claude Code 和 Codex 分别搭载了 30 多种工具，因为他们不得不这样做。他们的长尾需求比你的整个客户群还要大，因此他们必须支持各种用例。但这种对广度的妥协，导致他们的智能体上下文中充斥着额外的指令、工具和技能；而你可以砍掉所有冗余，专注于 2-3 个核心杀手级用例。举个例子，Pi 的精简架构在编码任务上击败了 Claude Code 和 Codex，这并不是因为 Anthropic 和 OpenAI 的人能力不行，而是因为 CC 和 Codex 服务着数百万客户，而 Pi 只服务于开发者和评估任务。
The goals of big labs are also often far more multidimensional than yours. Customers do not want a workflow that spawns 10 subagents, takes 20 minutes, and costs $30 for a quick PR review. This is also largely why Pi and other agent harnesses have their own niche in this coding harness ecosystem.大型实验室的目标通常也比你多元得多。客户并不想要一个为了简单的 PR 审查而衍生出 10 个子智能体、耗时 20 分钟且花费 30 美元的工作流。这在很大程度上也是为什么 Pi 和其他智能体架构能在当前的编码架构生态系统中拥有一席之地的原因。
You, unencumbered, should not and cannot lose a marathon against a guy wearing a weighted vest.作为一个无负担的参赛者，你不应该也不可能输给一个背着负重背心的对手。
You must sling models你必须灵活切换模型
By virtue of existing outside of any one lab's ecosystem, you are allowed to use any model you wish. This is your biggest structural advantage, as big as the sodden clay of Agincourt for the English. Do not cede it at any cost. Some recent anecdotes from Shortcut:由于身处任何单一实验室的生态系统之外，你可以自由使用任何你想要的模型。这是你最大的结构性优势，就像阿金库尔战役中英军面对法军那泥泞的土地一样。无论如何都不要放弃这一点。以下是 Shortcut 的一些近期案例：
- Shortcut's default has been Opus-family models for over a year. But when we benchmarked GPT 5.6 Sol, we found it was 2x cheaper and 2x faster than Opus 4.8 while . So we replaced Opus for Sol as the default model 24 hours after release.- Shortcut 一年多来默认使用 Opus 系列模型。但当我们对 GPT 5.6 Sol 进行基准测试时，发现它比 Opus 4.8 便宜 2 倍、速度快 2 倍，且准确率相当。因此，在 Sol 发布 24 小时后，我们就将其替换为默认模型。
- For reading images, PDFs, and scanned tables, we benchmarked the field of multimodal models through a combination of public and internal evals. Gemini Flash won by a clear margin on accuracy, not even considering cost. So Shortcut's image and PDF perception capabilities, which powers financial data extraction, runs on Gemini. Claude for Excel still has to read your image and PDF files with Claude models even when they are worse at it and costs a great deal more.- 在读取图像、PDF 和扫描表格方面，我们通过结合公开和内部评估，对各类多模态模型进行了基准测试。Gemini Flash 在准确率上以明显优势胜出，甚至还没考虑成本因素。因此，驱动金融数据提取的 Shortcut 图像和 PDF 感知功能现在运行在 Gemini 上。而 Claude for Excel 即使在效果更差且成本高昂的情况下，依然不得不使用 Claude 模型来处理你的图像和 PDF 文件。
- We recently benchmarked GLM 5.2 using our internal evals and found it was 2x slower, 2x cheaper, and scored 2-4% lower than SOTA models (79% for GLM 5.2 vs 81-83% for Opus48/Sol). This is an extremely attractive option for those that are cost-sensitive and do not need SOTA accuracy on the hardest Excel tasks. So no brainer, we serve GLM 5.2 routed through the same harness, same tools, same skills.- 我们最近使用内部评估对 GLM 5.2 进行了基准测试，发现它比 SOTA 模型慢 2 倍、便宜 2 倍，得分低 2-4%（GLM 5.2 为 79%，Opus 4.8/Sol 为 81-83%）。对于那些对成本敏感且不需要在最难的 Excel 任务上追求顶尖准确率的用户来说，这是一个极具吸引力的选择。所以毫无疑问，我们通过相同的架构、相同的工具和相同的技能来运行 GLM 5.2。
- Train your own model. We trained our own model, , off of Qwen3.5-27B, for finance and spreadsheet work, and it is in production. While it still has ways to go in accuracy and performance (~Sonnet 4.5 level), it is smaller, faster, and dirt cheap to serve. Even if it never gets to Fable level intelligence, the ability to offload tasks to worker subagents for spreadsheet tasks will optimize your cost-accuracy frontier. And in the long-term, building model training and serving expertise will allow you to own your own fate, especially for enterprise customers that require ZDR and customized deployments.- 训练你自己的模型。我们基于 Qwen3.5-27B 训练了自己的模型 Pivot，用于金融和电子表格工作，目前已投入生产。虽然它在准确率和性能上还有提升空间（大约在 Sonnet 4.5 水平），但它更小、更快，且服务成本极低。即使它永远达不到 Fable 级别的智能，将任务卸载给子智能体来处理电子表格任务，也能优化你的“成本-准确率”边界。从长远来看，建立模型训练和部署的专业能力将让你掌握自己的命运，尤其是对于那些需要 ZDR（零数据保留）和定制化部署的企业客户而言。
Use the full landscape of closed and open-source models to widen the gap between you and any big lab forced to operate within their walled garden.利用闭源和开源模型的全景生态，拉大你与那些被迫在“围墙花园”内运作的大型实验室之间的差距。
You care more你更在意用户
and I fly to New York to sit with customers and learn the workflows that actually matter to them. Not GTM, not FDEs, founders. Then our team holds their hands until each specific workflow works, one at a time. Then we turn those workflows into benchmarks and evals written in the spirit of the real tasks, and we hillclimb against them obsessively. Then we turn the research insights back into the product, so the accuracy we won for one customer gets delivered to all of them.@nicochristie 和我经常飞往纽约，与客户坐在一起，了解他们真正关心的工作流。不是市场推广，也不是销售工程师，而是创始人亲自上阵。我们团队会手把手地协助他们，直到每一个具体的工作流都能顺利运行。然后，我们将这些工作流转化为基于真实任务的基准测试和评估，并执着地进行迭代优化。接着，我们将研究洞察反馈到产品中，这样我们为一个客户赢得的准确率提升，就能惠及所有用户。
A lab likely cannot run this loop for your vertical, because your vertical is not their main quest. We can, because it's the whole business. This loop, of real workflows in, benchmarks and features out, is the magic sauce.实验室可能无法为你所在的垂直领域运行这种闭环，因为你的领域并非他们的主要任务。但我们可以，因为这就是我们的全部业务。这种“真实工作流输入，基准测试与功能输出”的闭环，就是我们的制胜法宝。
Expectations预期
Distribution is still the single biggest force in this market, but it doesn't win everything. It wins the enormous middle of the market that is genuinely fine with "good enough".分发渠道依然是这个市场上最强大的力量，但它并不能赢得一切。它能赢得的是那些对“足够好”感到满意的庞大中间市场。
But in every single domain, there is a real and valuable slice of users who demand the best, and will pick the product that is actually right more often even if it comes from a smaller company they had to go find. Being right more often is a thing you can win by providing model choice, a focused harness, and caring more. It's hard, it's exhausting, and it's also enough.但在每一个领域，都有一群真正有价值的用户，他们追求卓越，并且会选择那些更准确的产品，即使该产品来自一家他们需要费力寻找的小公司。通过提供模型选择、专注的架构以及更多的投入，你完全可以做到“更准确”。这很难，很累，但这也已经足够了。
想发布自己的文章？想发布自己的文章？
升级为 Premium升级为 Premium