Visual workflow builders are too complex

I agree with Harrison here. Workflow builders have their place and power many MANY capabilities at every enterprise. These have been a pain even back in the ETL days (Apache NiFi, anyone?). It's not that they don't work, they're just usually miserable to deal with at any appreciable scale. The thing that worries me the most is that ultimately these workflows are **code**. Which means they should be tested when changed, etc. Most non-technical folks don't realize what is really involved in building code, so issues quickly arise when they adopt these visual workflow tools. In the very near future, all we'll need is a Deep Agent + a dynamic MCP system. Send the AI your problem, give it the tools you want it to use, let it figure things out for you (and check in a long the way). No workflow needed.

Agreed! 💯 Visual Workflow builders like n8n or OpenAI's Agent Builder may seem very pretty and useful in small projects or POCs but not good enough for complex and production use cases.

Great piece, we def don't need another workflow builder. This UI is always ineffective for high-lvl tasks and even worse for novice learners who need better management skills at the early steps of AI-assisted work. Love that folks at LangChain see that aswell :3

Loved this thoughtful post by Harrison Chase on why the world doesn’t need another workflow builder. At Airtop, we believe English is the new programming language. You describe what you want, in plain English, and your agent figures out how to do it - navigating any site, logging in, fetching data, running tools. You don’t design flows; you just describe them. Beautiful post, Harrison - and a great reminder that we’re all moving toward the same goal: bringing the power of automation to everyone. And if you don’t believe it yet, try Airtop - build powerful web agents with just words.

I share Harrison ‘s perspective here. The market fit seems unclear, particularly given the limitation to OpenAI models only. Successful multi-agent workflows depend on leveraging the right model for each specific task. Restricting users to a single model provider fundamentally limits what’s possible. And if the aim is for non-technical users building agent workflows, the abstraction challenge runs deep…you need to hide complexity at multiple levels (JSON configurations, UI widgets, MCP, orchestration logic, etc.) while maintaining flexibility. What sets platforms like n8n , Make , Flowise ans others apart is their approach: they’re building ecosystems, not walled gardens. Their value comes from maximizing connectivity and model choice, not restricting it. Just my perspective, of course. 🙂

My view on this is that, a better UX for building agents is describing what an agent needs to do and having Claude Code/Cursor, etc, generate the code and deploy on serverless platforms to get someone a fully functional agent. Serverless Platforms have to be more than Lambda - agents need serverless compute with durable execution built into the runtime. That makes it easy for LLMs to generate code which implements the business logic without handling the distributed systems aspects of making the agent scale to infrastructure bleeps. Shout out to Harrison Chase for a thoughtful post.

Development/engineering efforts vs. low-code no-code? From the days of Data Science democratization to AI Agents, we face the same question. Microsoft, Dataiku, Databricks, and all the AI platforms are supported for a reason. Now, OpenAI has made a royal entry to the game with another Workflow Builder. After reading Phil Fersht (https://lnkd.in/g7k5ChCu), Harrison Chase, and Deedy Das (https://lnkd.in/gAn_KPg7) today, I believe there is a strategy building by OpenAI. I see Harrison Chase blog post analysing it very well for us. From my experience, low-code and no-code platforms always help. Utilize a no-code or low-code platform to prototype and validate quickly. Then rebuild properly for production. Where do you see the ceiling for no-code agents in enterprise? Which use cases will always need custom development?

Agreed! Observing AI products are mimicking Saas like workflows. Need simpler experiences that language models are very capable of supporting. We need to work backwards from the end customer experience if we truly want to accelerate AI adoption in enterprises. Sameer Goel Ramesh May

Agree with Harrison Chase that what he terms "No-Code Agents" (LLM prompt + tools) are surprisingly useful. I've been building many personal no-code agents for tasks from running stock option strategies to crypto trading to ecommerce. You can get started automating personal workflows simply using ChatGPT's little-known scheduled tasks feature, zero code required.

Agree with Harrison Chase! Two major things still needs to be addressed. How can we make it easier to create reliably good agents in a no-code way. These should be agents! Not low code workflows. How can we make code generation models better at writing LLM powered workflows/agents.