为了避免受到已经过度饱和的外部基准测试的偏见：https://arxiv.org/html/2602.16763v3，他们基于自己工程师团队经常执行的任务创建了自己的基准测试。结果符合我们的预期，但可能让业内许多人感到惊讶。用他们的话说，“…模型调用的框架显著影响成本和质量”，并且，“在许多情况下，像 Pi 这样简单的框架在我们的工作负载下表现最佳。”
当与 Opus 4.8, xhigh 结合使用时，Pi 的整体通过率最高，其成本远低于 Claude Code 和 Codex。
Pi 的出色之处在于它不会试图将模型包裹在一堆默认设置和指令中，而这些指令往往在指令层级中迷失：https://openai.com/index/the-instruction-hierarchy/。相反，Pi 不会干扰模型的运行，团队能够添加他们工作流程中真正需要的内容。
Databricks 的研究非常有见地，因为它将模型与外壳分离。
他们报告称，当他们通过不同的外壳运行相同模型并进行相同的思考努力时，“每个任务的成本差异显著（有些情况下超过两倍），而质量保持不变”。我们称之为 Pi 的“上下文纪律”。“Pi 每轮发送的上下文量大约减少了 3 倍。它更好地管理上下文，保持更紧凑的工作集，并在更少的运行次数内完成任务。”
前沿模型现在通常非常擅长理解终端（或终端风格）的编码环境，并在其中执行操作。Anthropic 最近将 Claude Code 的系统提示减少了 80%：https://x.com/petergyang/status/2078895219534438556?s=20，这就是一个明显的例子。因此，问题不再是工具的原生性，而是它如何处理上下文以避免冗余并使用干净的原语进行操作。模型需要一个干净的环境接口，以及一个不会浪费上下文的工具。
Pi 提供了这一点：减少提示开销和重复上下文，降低运行成本，减少不必要的抽象。因为它是可扩展的，您不会失去能力，而是获得选择性。只有当复杂性“值得它的价值”时，您才会增加它。
我们还看到本地模型发展迅速，在 Earendil：https://earendil.com 上我们发现它们非常有前景。Pi 的上下文规范在这里尤其是一个优势。本地模型通常上下文窗口较小，而预填充：/posts/prompt-caching/ 可能需要很长时间，因此保持稳定的提示前缀非常重要。上下文规范意味着我们不会在用户未明确要求时更改上下文，从而避免长时间的重新预填充。结合最小化的默认系统提示和工具集，这使得 Pi 成为本地模型的理想工具。
Pi 证明它可以管理所有这一切。以更低成本、更简约、更高性能的方式运行。
Subject: Pi, Minimal and Performant
AI has made code cheap, and as a result many companies are building bigger tools in pursuit of better performance. Larger prompts, more orchestration, more layers, more complexity. This also makes these tools intrinsically more expensive to use. Pi takes the opposite approach.
Pi is the coding harness that chooses minimalism on purpose. It comes out of the box with only 4 tools, and its system prompt：https://github.com/earendil-works/pi/blob/main/packages/coding-agent/src/core/system-prompt.ts#L121-L159 and tool definitions come in below 1,000 tokens. The idea being that most work can be done with the basics, and if you want more, build it.
Evidence increasingly suggests that Pi’s design is not just cleaner; it’s cheaper and more performant. Users are finding that vanilla Pi produces industry leading results, even before adding on extensions to match user specific workflows and needs. As we'll see in case studies of Databricks and Shopify, Pi produced ideal outcomes for both.
Databricks recently shared their findings “ Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase ：https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase.” The goal of their research was to understand which coding agents offer the best performance on real-world coding tasks, and how task-performance varies with price.
To avoid bias from external benchmarks that have become oversaturated：https://arxiv.org/html/2602.16763v3, they created their own based on tasks their team of engineers regularly performs. The results match what we would expect, but what many in the industry may have been surprised to learn. In their words, “...the harness a model is called from dramatically impacts cost and quality,” and, “in many cases, simple harnesses like Pi performed best on our workloads.”
When combined with Opus 4.8, xhigh, Pi had the highest overall pass-rate, at a significantly lower cost than both Claude Code and Codex.
Pi shines because it doesn’t try to wrap the model in a bunch of defaults and instructions that get lost in the instruction hierarchy：https://openai.com/index/the-instruction-hierarchy/. Instead, Pi stays out of the model’s way, and the team is able to add what they actually need for their workflow.
Databricks’ study is insightful because it separates model from harness.
They reported that when they ran the same model with the same thinking effort through different harnesses, “the cost per task differed significantly (more than 2x in some cases), while quality remained the same”. We call this Pi’s “context discipline”. “Pi sent about 3x less context per turn. It managed context better, keeping a tighter working set and finishing the tasks in fewer runs.”
We agree that one must take into account end-to-end engineering economics, and not just price per token. And this is also true at the model level; we have observed, for instance, that running complex workflows on Haiku 4.5 was often more expensive than Sonnet 4.6, especially when code execution was involved, simply because the agent required more turns to complete the task successfully.
Now we see this at the harness level too; stronger, more expensive models with a performant harness can be cheaper than the converse.
Minimalism is part of Pi’s core philosophy. What makes this work is that minimal does not mean inflexible. In fact, it is the first widely used agentic infrastructure created for extensibility and self-editability.
Another insightful external validation of Pi’s design comes from Shopify. In this post from Shopify Engineering：https://shopify.engineering/autoresearch, David Cortés describes building pi-autoresearch directly as a Pi extension, by simply asking “Pi, [to] create an extension for Autoresearch...”. Pi reads its own extension documentation and starts building a new workflow from there.
Autoresearch is an autonomous loop for optimization with coding agents. When you ask for a change, it runs experiments to find out what works and what causes regressions. For as long as the target is measurable, it can throw out these regressions and keep self-improving.
For Shopify and others：https://x.com/pidotdev/status/2080616483072225778?s=20, the Autoresearch extension quickly became a serious internal productivity tool. Shopify reported cases including unit tests running “300 times faster,” React component mounting “20% faster,” reduced build times across multiple projects, and even improvements to pnpm performance.
The important point here is that Pi doesn’t ship any of these tools out of the box. Instead, it makes it ridiculously simple for you to build them. Instead of assuming the vendor knows your workflow and trying to ship every tool under the sun, Pi assumes you know best, and gifts you extensibility to wield and craft your own workflow.
About a year ago, an argument could be made for native harnesses having a structural advantage over all others, because models were built around them. However, this argument has gotten weaker.
Frontier models are now generally very competent at understanding a terminal (or terminal-style) coding environment, and acting within it. Anthropic recently cutting down Claude Code’s system prompt by 80%：https://x.com/petergyang/status/2078895219534438556?s=20 is a clear sign of this. So the question is becoming less about how native the harness is, and more about how it handles context to avoid redundancy and act with clean primitives. Models need a clean interface to the environment, and a harness that does not waste context.
Pi provides this: less prompt overhead and repeated context, cheaper runs, fewer unnecessary abstractions. Because it is extensible, you do not lose power, but gain selectivity. You add complexity only when it “earns its keep”.
We are also seeing local models developing fast, and at Earendil：https://earendil.com we find them very promising. Pi’s context discipline is especially an asset here. Local models usually have lower context windows, and prefill：/posts/prompt-caching/ can take a long time, so preserving a stable prompt prefix matters. Context discipline means we do not change the context without the user explicitly asking for it, avoiding minute-long re-prefilling. Combined with the minimal default system prompt and tool set, this makes pi an ideal harness for local models.
Pi is proving that it can manage it all. To be cheaper, minimal, and more performant.
情报判断
Aioga 编辑摘要
Pi 刻意采用极简设计，开箱仅提供 4 个工具，系统提示词与工具定义合计低于 1,000 tokens。材料称，Databricks 测试中，Pi 搭配 Opus 4.8（xhigh）取得最高整体通过率。
对编码代理的采购与技术选型而言，值得关注的不只是底层模型，还包括框架如何组织上下文、工具和指令。材料称 Pi 每轮发送的上下文约为其他工具的三分之一，这可能成为控制成本的重要因素。 建议继续核对 Databricks 基准的任务构成、通过率口径、成本计算方式及各代理配置，并在自身代码库上进行同条件测试。现有材料不足以证明 Pi 在所有项目、模型和工作流中都保持领先。
扫码直达文章详情