Anthropic released a new product today.

The positioning is: You don’t need to do Harness, I will do it for you. I know the limitations of the model best. I will design Harness according to the characteristics of the model and package it for you. And this package is Managed Agents.

This scene is very similar to ten years ago, when AWS first educated the market on what cloud computing was. When everyone thought that "going to the cloud" was necessary, AWS said: If you can't build it yourself, I will do it. Most businesses end up choosing hosting. Because infrastructure has never been the core competitiveness.

Anthropic does exactly the same thing.

I will post three blogs to teach the entire industry step by step what a harness is, why a harness is more important than a model, and how to design a good harness. (Anthropic said: Don’t wait for the next generation model, make Harness immediately! Anthropic said: The Harness idea spread online is outdated, just do these three things! Anthropic product leader: PRD is dead, long live the prototype)

Using context anxiety as an example, we launched a new product.

Sonnet 4.5 will become "anxious" when it reaches the limit of the context window, stop work early, and suffer a sharp decline in quality. So they early added sprint + context reset to the harness to deal with it.

After Opus 4.5 came out, the anxiety disappeared. Sprint was killed directly. Previous patches became a liability.

So, the previous conclusion is: Harness's possibility space is not shrinking, it is just moving. With each generation of models, we have to re-examine which constraints are still effective.

But now look at this example from another angle:If you build that set of sprint + context reset yourself, you will waste engineering time. If Anthropic manages the Harness for you, they will update the Harness when they upgrade the model. You don't have to move anything.

This is the main selling point of Managed Agents:The shelf life of the harness is so short that it is not cost-effective for you to maintain it yourself.

How is it technically done?

The previous Claude Agent SDK was equivalent to putting inference loop, code execution, session recording... all in one container. The container hangs and the session is lost.

Now they have broken it into three pieces:

• Brain: Claude and his harness, responsible for thinking and decision-making

• Hand: sandbox and tools, responsible for execution

• Memory: independent session log, records everything

These three are independent of each other. The container is down? New containers can be rebuilt. Harness hangs, the new harness gets the session log and continues from the breakpoint. Nothing is irreplaceable.

Of course the core point is: Harness has become a replaceable hot-swappable module. The one who has the power to hot-swap is Anthropic, not you.

Master Ma Dongxi has a good summary:

Harness doesn't care whether the sandbox is a container, a phone, or a Pokémon emulator. As long as it conforms to the interface of "name and input go in, string comes out".

Write at the end

Before, everyone was saying: Harness means that the truly scarce capabilities are not inside the model, but outside the model.

But today Anthropic is trying to put the "outside of the model" layer back into the "inside of the model manufacturer".

Harness Engineering is not dying. But the window period for "people making harnesses" may be shorter than everyone expects.

Your Harness will either be eliminated by model advancement, replaced by platform services, or you will have to run faster than both.

However, from a realistic perspective, general Harness can cover most scenarios, but not all scenarios. Agents in each field, including law, finance, and medical care, have their own evaluation standards and security boundaries. It is not difficult to achieve 80 points for a general platform, but the last 20 points are the competitiveness.

What's more, the open source community will also take many more interesting routes, such as Hermes shared yesterday.

Let the Agent write the Skill himself, the Agent uses the Skill, and the Agent changes the Skill. Instead of handing Harness to the platform, let the Agent write its own Harness.