🔶 AI 工作流构建器之死
英文标题：🔶 The death of AI workflow builders
曾被视为 AI Agent 开发未来的可视化无代码工作流编辑器，在不到一年内迅速失势。随着 Claude Opus 4.5、Claude Code 和 Codex 等模型及编程 Agent 的能力提升，开发者可以更低成本地生成和维护确定性的工作流代码，而非在复杂 UI 中配置节点图。文章认为，未来属于由稳健 Agent harness 支撑的确定性脚本：Agent 在脚本失败时介入诊断、修复和测试，但大规模生产运行、维护以及保证可验证正确的输出，仍是重要挑战。
38.2 KB
32.8 KB
98.4 KB
672 B
中文
AI 工作流构建器之死
可视化无代码编辑器曾看起来是 AI Agent 的未来。然而不到一年后，编程 Agent 接管了一切。
当 OpenAI 于 2025 年 10 月发布 Agent Builder 时，所有人都预测，它将成为未来构建 Agent 的方式。
团队不再需要编写编排代码，而是可以在画布上拖放 Agent 节点、工具、护栏和分支，再将它们连接起来。
大约在同一时期，Microsoft 在 Copilot Studio 中推出了可视化 Agent 流程，Google 也公布了 Vertex Agent Designer。这些界面的外观都颇为相似，通常基于热门的开源无代码编辑器 React Flow 构建。
快进到今天，它们全都已经被弃用或彻底关闭。
在过去几周里，我还注意到，多家打造 AI 工作流构建器的初创公司已经关闭或被人才收购（acquihire）。
在当时，这些 AI 工作流构建器看起来确实很有吸引力：
- 上下文窗口容量有限。
- 工具调用和 Agent 循环不可靠且成本高昂。我记得曾在一个陷入无限循环的工作流上烧掉海量 token。
- 更复杂的任务需要手动将多个提示词串联起来。当时还没有 harness engineering 这个概念。
- 能够拖放 AI 节点的可视化画布，让这种复杂性变得更易于管理。
大约在 OpenAI 发布 Agent Builder 一个月后，Claude Opus 4.5 发布，Claude Code 开始流行，Codex 也很快跟进。
模型在规划、工具调用以及遵循长指令方面变得更强。编程 Agent 已能够自行创建和维护确定性的工作流代码；与其在 UI 中配置复杂的可视化图，这种方式容易得多。
因此，随着上述暂时性的模型局限消失，AI 工作流构建器的需求也随之消退。Flowise 的停止运营公告提到：“开发者已经转向编程 Agent。”
我认为，未来属于确定性的工作流脚本，并在其周围配备一个健壮的 Agent harness：当脚本失败并需要适应变化时，harness 会介入。我们在 Kadoa 使用这种方式构建具备自我修复能力的网页数据管道：当网站或 PDF 的版式发生变化时，Agent 会调查问题、修复并测试提取或转换代码。如果它无法解决，则将问题升级给人工操作员。
生成确定性的工作流代码已经变得非常便宜；但要在生产环境中大规模运行和维护这些代码，成本并未降低。确保输出结果可被证明为正确，仍然是一项重大挑战。
English / 原文
The death of AI workflow builders
The visual no-code editors looked like the future of AI agents. Less than a year later, coding agents took over.
When OpenAI launched their Agent Builder in October 2025, everyone predicted it was the future of how we’d build agents.
Instead of writing orchestration code, teams would drag-and-drop agentic nodes, tools, guardrails, and branches onto a canvas and connect them.
Around the same time, Microsoft launched visual agent flows in Copilot Studio and Google announced Vertex Agent Designer. The interfaces all looked kind of similar and were usually built on the popular OS no-code editor React Flow.
Fast forward to today, all of them are deprecated or shut down completely.
Over the last few weeks, I also noticed that several startups building AI workflow builders have shut down or got acquihired:
At the time, these AI workflow builders looked compelling:
- Context windows were limited.
- Tool use and agent loops were unreliable and expensive. I remember burning a gazillion tokens on a workflow stuck in an infinite loop.
- More complex tasks required manually chaining prompts together. Harness engineering wasn’t a thing yet
- A visual canvas where you could drag-and-drop AI nodes made that complexity manageable
About a month after OpenAI launched their Agent Builder, Claude Opus 4.5 got released and Claude Code took off, with Codex following soon after.
Models became better at planning, tool use, and following long instructions. Coding agents became capable of creating and maintaining deterministic workflow code themselves, which became much easier than configuring a complicated visual graph in a UI.
So as the above temporary model limitations disappeared, so did the need for AI workflow builders. Flowise’s sunset mentions that “developers moved on to coding agents”.
I think the future lies in deterministic workflow scripts with a robust agent harness around them that jumps in whenever the script fails and needs to adapt. We use this at Kadoa to build self-healing web data pipelines: when a website or PDF layout changes, an agent investigates, fixes, and tests the extraction or transformation code. If it can’t figure it out, it escalates to a human operator.
Generating deterministic workflow code got very cheap, running and maintaining it at scale in production did not. Ensuring provably correct output is still a big challenge.