本文探讨了AI工程师社区中关于构建智能体的两种主要方法：基于大模型的方案和基于工作流的方案。文章讨论了OpenAI的“构建智能体实用指南”引发的争议，并分析了这两种方法各自的优缺点。文章指出，大模型方案在模型更新时具有优势，而工作流方案则更容易修改代码。最终，作者认为这两种方法并非对立，而是在构建AI智能体时应该同时考虑的选择，并强调了智能体框架的重要性。

🧠 **大模型方案的优势**：大模型方案在面对模型更新时更具优势，工程师无需投入大量时间调整工作流。OpenAI和Gemini的研究成果表明，通过通用智能体增强模型，可以减少对工作流工程的需求。

🛠️ **工作流方案的优势**：工作流方案更容易修改代码，工程师可以更灵活地调整和优化系统。这种方法允许从一个方案逐步过渡到另一个方案，以优化代码的可修改性。

🤝 **两种方案并非对立**：作者认为，在构建AI智能体时，大模型方案和工作流方案并非对立关系，而是应该同时考虑。Harrison Chase也持相似观点，认为应该为用户提供两种选择。

📊 **智能体框架的重要性**：文章强调了智能体框架的重要性，并提到了一个比较了各种智能体框架的表格，该表格涵盖了意图、记忆、规划、授权、控制流和工具等关键特性。这有助于AI工程师评估和选择合适的框架。

Quick reminder: AI Engineer CFPs close soon! Take a look at “undervalued tracks” like Computer Use, Voice, and Reasoning, and apply via our CFP MCP (talks OR workshops, we’ll figure it out).

Relevant to today’s quick post we do have an Agent Reliability track. Also: take the 2025 State of AI Engineering Survey!

The AI attention economy has enabled a hypeboi priesthood who exist in a perpetual state of performative orgasmic nirvana, minds continually blown as every launch Changes Everything, vibing at gigahertz oscillations of “it’s so over” vs “so back”. OpenAI’s “Practical Guide to Building Agents” is the latest such earth shatterer:

This guide, however, has been less well received than Anthropic’s equivalent.

If you watch his multiple appearances with us, Harrison Chase is not someone who is quick to “anger”, so calling this guide “misguided” and doing a word by word teardown can seem like fighting words for him1.

At the heart of the battle is a core tension we’ve discussed several times on the pod - team “Big Model take the wheel” vs team “nooooo we need to write code” (what used to be called chains, now it seems the term “workflows” has won).

Team Big Workflows

You should read Harrison’s full rebuttal for the argument, but minus the LangGraph specific parts, the argument that stood out best to me was that you can replace every LLM call in a workflow with an agent and still have an agentic system:

And the ideal agent framework lets you start from one side of the spectrum and move to the other, optimizing for making code easy to change:

Team Big Model

To be clear it’s easy to understand where the Big Model folks are coming from: if you work with Big Lab enough, you’ve seen hundreds of engineer-hours of hand tuned workflows obliterated overnight with the next big model update — the AI Engineer equivalent of learning the Bitter Lesson again and again. This is why “AI engineering with the Bitter Lesson in mind” was such a resonant topic at the Summit (now at 124k views across platforms):

Specifically, I think the success of both OpenAI and Gemini’s Deep Research this year primarily leveraging O3 to reason through research planning and execution, and later Bolt and Manus AI doing the same with Claude, with very little workflow engineering, has demonstrated that there’s a lot to be said for building general purpose agents that simply augment models without the “inductive bias” constraints of workflows. O-team researcher Hyung Won Chung noted that adding more structure gets you wins in the short term, but that structure tends to lose in performance as the model (pretrain or inference compute) keeps scaling up.

If your goal is to build AGI, to build a horizontal platform, particularly one targeting non-technical consumers who are confused by even having a model selector, then it’s an understandable position to take, and (even encourage, for the purposes of dataset/human feedback collection).

Workflows AND Agents, not OR

Ultimately the reason I argue Harrison isn’t -actually- taking a fighting stance is he leaves room for the spectrum to exist and makes a remarkably (for someone with obvious skin in the game) balanced argument that you’re going to just want options for doing both:

I find this hard to debate - if meaningful conversation is to be had, it really is more about where the current state of this Pareto frontier really is today (I’m not sure it is convex yet) and how to move it out.

IMPACT of Agent Frameworks

The other pretty cool thing that Harrison did in his piece was publish a full comparison table of all relevant Agent Frameworks today, although of course even he couldn’t escape the McCormick trap. It’s useful to test our descriptive framework of everybody’s Agent definitions against a new out-of-distribution Agents definition:

I think this is a remarkably fair shopping list of abstractions and features for the discerning Agent Engineer — it also articulates why you feel certain gaps when an Agent Framework promises you the world and yet you can’t do some things easily.

The Great Debates

If you like this kind of debate, we’re doubling down on the success of the Dylan v Frankle showdown from last year’s NeurIPS, and also accepting submissions for what we’re calling “The Great Debates” - good faith debaters from two sides of a relevant industry debate. Everybody wins, but the people who are best able to change minds win the most. Apply in pairs!