OpenAI scientists suggest: Don’t put too much effort into Harness—next-generation models may be built-in.

OpenAI research scientist Noam Brown focuses on developing reasoning models. He has publicly advised developers not to put too much effort into complex Agent frameworks (Harness), arguing that model progress is too fast: the functionalities that the framework produces today may become native capabilities of the model itself within a few months. Alexander Embiricos, OpenAI’s head of enterprise products, also echoed this stance.
(Background: A quick overview of the four AI Agent tracks—frameworks, launchpads, applications, and memes)
(Additional context: OpenAI released its strongest reasoning models, o3 and o4-mini—able to think about images, automatically select tools, and achieve further breakthroughs in math and coding performance.)

Table of Contents

Toggle

• Save your energy—let the model handle it
• The other camp doesn’t agree: frameworks are the true moat
• The controversy is a wager against the next model version

Key Takeaways

• Noam Brown advises against pouring effort into a complex Agent framework
• Noam Brown leads the development of the o-series reasoning models
• Opposing experiments show that modifying the framework in one afternoon boosted the coding performance of 15 models across the board

While the entire AI community is busy making Agent frameworks thicker and thicker, people at OpenAI say not to build them. OpenAI senior researcher Noam Brown publicly advised developers not to put too much energy into complex Agent frameworks (Harness)—the pipeline wrapped around the model, responsible for connecting tool calls and breaking down steps, but not doing real thinking itself. Since model capabilities are improving too quickly, the features that are squeezed out through the framework today are very likely to become native capabilities of the model within the next few months. He suggests keeping the framework simple and having the model itself handle more of the work.

The Noam Brown who urged everyone not to build frameworks is one of the key drivers behind OpenAI’s reasoning model lineup (the o-series, e.g., o1, o3). He mainly works on test-time compute—making models spend more computation on thinking. Before joining OpenAI, he worked at Meta, where he built well-known poker AI and strategy game AI. That background gives his views extra weight when it comes to model capabilities.

Noam Brown has cited an example before reasoning models emerged: developers had to rely on lots of engineering to make non-reasoning models like GPT-4 perform complex decomposition, repeatedly call things, break tasks into steps, and force reasoning behavior through external orchestration. Once o1 was released, almost all of that engineering became essentially meaningless. The carefully constructed external add-on framework actually made results worse. Simply throwing the problem at the reasoning model—without any scaffolding—produced better results.

Save your energy—let the model handle it

Alexander Embiricos, OpenAI’s head of enterprise products, also echoed the same position. He said the company would deliberately avoid manually developing capabilities that future models should be able to have natively themselves. Features that engineers stay up late building today may just be doing contract work for the next version of the model—and are doomed to be replaced before they even go live. OpenAI’s own Codex team put it even more plainly: “Scaffolding is propping things up, not expanding capability.”

The other camp doesn’t agree: frameworks are the true moat

This is actually a debate that’s been unfolding— not everyone is buying into it. Jerry Liu, founder of LlamaIndex, has said the opposite: “The framework is everything.” He believes that context engineering (the craft of feeding context into the model) and workflow design are the key to blocking developers from extracting AI value.

The opposing side’s evidence is that in February 2026, in just one afternoon, only changing the framework—without switching models—made the code-writing performance of 15 large language models improve dramatically across the board. Others have also observed that nearly every Agent that goes live eventually converges to the same core loop: call a tool, get the result, put it into the context, then ask the model again. The framework architecture itself may be the core value of the product. Both camps can show real results, and no one has truly won this battle yet.

The controversy is a wager against the next model version

At its core, this debate is about the same question: whether to race against model progress. It brings to mind the long-circulating “The Bitter Lesson” in AI: general methods powered by compute often win over human-crafted cleverness over the long run. Putting your moat on framework engineering, to some extent, is placing a wager against the next model version—betting that it won’t learn the functionality you painstakingly built.

This reminder is not just something the crypto industry can watch from across the river. AI Agents are one of the hottest narratives of the past two years, and many teams have put their moats on the manually built, complex Agent frameworks. Noam Brown’s message is right in front of you: the framework you build today might be eaten by the model’s native capabilities in a few months. As for which functions the next model will absorb—no one can guarantee, including Noam Brown himself.

FAQ

What is an AI Agent Harness (framework)?

Harness is the pipeline wrapped around the model that handles tool calling and step orchestration, without doing real thinking itself. For example, the layer that connects calls and retries is the Harness.

Why does Noam Brown advise developers not to overinvest in Agent frameworks?

Because model capabilities are improving too quickly: the features produced by using frameworks today may very likely become native capabilities of the model within the next few months, so overinvesting will likely be a waste of effort.