什么是 harness 工程？Google 用 ADK 2.0 与 Antigravity SDK 演示自动修复编码循环

如何在不手动编写任何代码的情况下交付一个软件产品？

今天一个朋友问我这个问题，我意识到自己没有一个简单的答案。所以我进行了深入研究。

事实证明，答案在于你如何设计你的包裹系统。

等等，什么？什么是包裹系统工程？

这也是为什么这成为目前关于编码代理最重要的趋势的原因。如今最大的问题是如何在不阅读每一行的情况下验证AI生成的代码。你如何确保代理不会破坏生产环境或删除你的数据？

OpenAI的一篇博客：https://openai.com/index/harness-engineering/ 分享了一个有趣的实验，一个由3名工程师组成的团队已经构建并交付了一个内部测试软件产品，完全没有手动编写代码。每一行代码——应用逻辑、测试、CI配置、文档、可观测性以及内部工具——都是由Codex编写的。

他们是如何做到的？他们没有编写应用程序。他们设计了包裹系统。

把AI代理想象成一匹强大的赛马。包裹系统就是那个赛道、眼罩和骑手的缰绳，它可以让马朝正确方向奔跑，而不是跳进看台。

正如我的同事Arthur Thompson今天解释的：对于代理来说——包裹系统由所有封装大语言模型（LLM）的确定性组件组成。

Balaji Subramaniam在他的博客中详细说明了这些确定性组件：https://medium.com/google-cloud/harness-engineering-for-multi-agent-systems-using-google-adk-2-0-e248b885cb95 ——包括编排层、执行沙箱、状态持久化和验证工具。

如果你想构建可靠的代理系统，你的工作将从编写逻辑转向设计环境。你需要关注以下内容：

这在实践中是什么样子？这里是一个使用谷歌反重力SDK的简单例子：https://antigravity.google/product/antigravity-sdk?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog，结合谷歌的ADK：https://adk.dev/2.0/ 来配置本地包裹系统。注意我们是如何严格将代理限制在特定工作区（workspaces=["./sandbox"]）并为它提供保存记忆的地方（save_dir="./trajectories"）从而让它能够从以往经验中学习的：

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F706wjvl4dxy8dnxgkbib.png

有了这个设计，你可以将遗留代码放入沙箱，编写一个简单的循环来运行单元测试，并让代理迭代地修复自身的错误。

那么，我们实际上如何对沙箱代理运行测试呢？

在现代测试框架工程中，测试是代理工作流图中的一个主动部分。使用 Google 的 ADK 2.0：https://adk.dev/2.0/，它引入了基于图的工作流，你可以将测试验证步骤定义为一个简单的路由节点。

如果测试通过，工作就完成了。如果测试失败，测试框架会自动将错误循环反馈给代理再次尝试。注意内置的“终止开关”：我们跟踪迭代次数，如果代理陷入不断出错和修复代码的无限循环，测试框架会安全地终止运行。

如果你想看到这个测试路由模式的实际效果，可以查看 Balaji 的 ADK 测试框架仓库中的完整实现示例：https://github.com/balajismaniam/adk-harness-engineering/blob/main/workflows/workflows.py。

要连接代理和测试节点，你可以使用工作流图来精确映射执行流程，而无需复杂的嵌套 Python while 循环。

把它想象成给赛道画出实际的跑道车道：

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fe6gyazb82hdpd8wqq4dv.png

恭喜！你已经构建了一个自主系统。代理编写代码并交给测试节点。如果测试失败并返回 loop_back 路由，代理会在手握错误日志的情况下再次尝试。

在 ADK 示例中查看更多循环模式的例子：https://github.com/google/adk-python/tree/main/contributing/samples/workflows/loop

你可能会想，为什么运行一个代理需要 Python 脚本。在普通的聊天窗口中，你就是那个管理工具：你会复制错误日志并照看模型。软件管理工具可以让系统自己照看自己，从而让你能够完全自动化测试驱动的编码，或者安全地重构庞大的遗留代码库。

要在你自己的机器上运行这个自我修复循环，设置时间不到五分钟：

从这里开始，你可以将我们简单的测试节点替换为一个实际执行 pytest 或 npm test 的子进程来检测你的沙盒，这样你就会拥有一个完全可用的修复循环。

如果你已经准备好扩展，可以下载完整的 IDE 和 CLI，地址是 antigravity.google：https://antigravity.google/?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog，探索 Antigravity 管理的代理：https://ai.google.dev/gemini-api/docs/antigravity-agent?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog 以进行远程执行，以及使用基于图的工作流的 Google ADK 2.0：https://adk.dev/2.0/。

我在 Google 的同事们整理了一些关于下一步该去哪里的一些很棒的指南。想了解如何为你的代理构建安全环境，请查看 Sara 展示的 Cloud Run 沙盒 codelab：https://codelabs.developers.google.com/codelabs/cloud-run/cloud-run-personal-agent-coffee-shop?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog。如果你想精通自我纠正，Balaji Subramaniam 最近发布了一篇深入探讨编码代理循环工程的文章：https://medium.com/@BalajiBuilds/61c30c9e36ca。要查看这些应用于大型企业用例的情况，请阅读 James O'Reilly 关于使用代理流水线和 Antigravity 自动化遗留系统现代化的大型案例分析：https://codelabs.developers.google.com/automating-modernization-with-antigravity?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog。

模板可以帮助你快速回答常见问题或存储可重复使用的代码片段。

你确定要隐藏此评论吗？它将在你的帖子中变为隐藏状态，但仍可通过评论的永久链接查看：#。

进一步操作中，你可以考虑屏蔽此人和/或举报滥用行为：/report-abuse

Google AI Studio 是开始使用 Gemini 构建的最快方式。准备好构建了吗？

DEV 社区：/ — 一个讨论、跟进软件开发并管理你的软件职业的空间

建立在 Forem：https://www.forem.com 上 — 开源软件：https://dev.to/t/opensource 支持 DEV：https://dev.to 及其他包容性社区。

用爱和 Ruby on Rails 制作：https://dev.to/t/rails。DEV 社区 © 2016 - 2026。

我们是一个程序员分享、保持最新状态并发展职业的地方。

How do you ship a software product with 0 lines of manually-written code?

A friend asked me this today, and I realized I didn't have a simple answer. So I dug deeper.

It turns out the answer is in how you engineer your harness.

Wait now, what? What is harness engineering?

There is a reason this is the most important trend right now around coding agents. The biggest question these days is how to validate AI-generated code without reading every single line. How do you make sure an agent doesn't break production or delete your data?

A blog by OpenAI：https://openai.com/index/harness-engineering/ shared an interesting experiment where a team of 3 engineers have built and shipped an internal beta of a software product with 0 lines of manually-written code. Every line of code: application logic, tests, CI configuration, documentation, observability, and internal tooling, has been written by Codex.

How did they do it? They didn't write the app. They designed the harness.

Think of an AI agent like a powerful racehorse. The harness is the track, the blinders, and the jockey's reins that keep it running in the right direction instead of jumping into the stands.

As my colleague Arthur Thompson explained today: for agents — the harness is composed of all the deterministic components that wrap the LLM.

Balaji Subramaniam details those deterministic components in his blog：https://medium.com/google-cloud/harness-engineering-for-multi-agent-systems-using-google-adk-2-0-e248b885cb95 — the orchestration layer, execution sandboxing, state persistence, and verification tools.

If you want to build reliable agentic systems, your job shifts from writing the logic to designing the environment. Here is what you need to focus on:

What does this look like in practice? Here is a simple example using the Google Antigravity SDK：https://antigravity.google/product/antigravity-sdk?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog with Google's ADK：https://adk.dev/2.0/ to configure a local harness. Notice how we are strictly bounding the agent to a specific workspace (workspaces=["./sandbox"]) and giving it a place to save its memory (save_dir="./trajectories" ) so it can learn from previous experience:

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F706wjvl4dxy8dnxgkbib.png

With this design in place, you can drop your legacy code into the sandbox, write a simple loop to run unit tests against it, and let the agent iteratively fix its own bugs.

So, how do we actually run tests against this sandboxed agent?

In modern harness engineering, tests are an active part of the agent's workflow graph. Using Google's ADK 2.0：https://adk.dev/2.0/, which introduces graph-based workflows, you can define a test validation step as a simple routing node.

If the test passes, the job is done. If it fails, the harness automatically loops the error back to the agent to try again. Notice the built-in 'kill switch': we track the iteration count so if the agent gets stuck in an infinite loop of breaking and fixing code, the harness safely pulls the plug.

If you want to see this test routing pattern in action, you can check out an example with a full implementation in Balaji's ADK harness repository ：https://github.com/balajismaniam/adk-harness-engineering/blob/main/workflows/workflows.py.

To connect the agent and the test node, you can use a Workflow graph to map out exactly how the execution should flow without needing complex, nested Python while loops.

Think of this as drawing the actual lanes on the racetrack:

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fe6gyazb82hdpd8wqq4dv.png

Congratulations! you've built an autonomous system. The agent writes the code and hands it off to the test node. If the test fails and returns a loop_back route, the agent tries again with the error log in hand.

See more examples of loop patterns in ADK samples. ：https://github.com/google/adk-python/tree/main/contributing/samples/workflows/loop

You might wonder why you need a Python script to run an agent. In a normal chat window, you are the harness: you copy the error logs and babysit the model. A software harness lets the system babysit itself, allowing you to fully automate test-driven coding or safely refactor massive legacy codebases.

To run this self-healing loop on your own machine today, the setup takes less than five minutes:

From there, you can swap out our simple test node for a subprocess that actually executes pytest or npm test against your sandbox, and you will have a fully functioning repair loop.

If you are ready to scale this up, you can download the full IDE and CLI at antigravity.google：https://antigravity.google/?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog, explore the Antigravity managed agent：https://ai.google.dev/gemini-api/docs/antigravity-agent?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog for remote execution and google's ADK 2.0：https://adk.dev/2.0/ for using graph based workflows.

My colleagues at Google have put together some incredible guides on where to go next. To learn how to build secure environments for your agents, check out Sara's codelab showcasing Cloud Run sandboxes：https://codelabs.developers.google.com/codelabs/cloud-run/cloud-run-personal-agent-coffee-shop?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog. If you want to master self-correction, Balaji Subramaniam recently published a deep dive on Loop Engineering for Coding Agents：https://medium.com/@BalajiBuilds/61c30c9e36ca. And to see all of this applied to a massive enterprise use case, read James O'Reilly's breakdown of Automating legacy modernization at scale using agentic pipelines and Antigravity：https://codelabs.developers.google.com/automating-modernization-with-antigravity?utm_campaign=CDR_0x91b1edb5_default_b550513795&utm_medium=external&utm_source=blog.

Templates let you quickly answer FAQs or store snippets for re-use.

Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink：#.

For further actions, you may consider blocking this person and/or reporting abuse：/report-abuse

Google AI Studio is the fastest way to start building with Gemini. Ready to build?

DEV Community：/ — A space to discuss and keep up software development and manage your software career

Built on Forem：https://www.forem.com — the open source：https://dev.to/t/opensource software that powers DEV：https://dev.to and other inclusive communities.

Made with love and Ruby on Rails：https://dev.to/t/rails. DEV Community © 2016 - 2026.

We're a place where coders share, stay up-to-date and grow their careers.