Anthropic personally enters the arena, and another batch of Agent startup companies have died.

Whenever a large model vendor unveils a new capability, a familiar stench of doom wafts through the AI startup circle.

This time, Anthropic has sharpened the blade itself.

On April 8th, Anthropic released Claude Managed Agents.

This isn't a new model but a set of operating layers. It turns the infrastructure that developers previously had to build and maintain themselves into a platform capability.

The official documentation states clearly: This is a fully managed agent harness. Developers can create agents, configure containers, and run sessions through the API.

Put simply: Anthropic has started selling the Agent operating layer.

Entrepreneurs know best what this means.

Over the past year, they've worked around the clock with investors' money to fill the gap in the platform capability that Anthropic is now offering.

But now, this batch of Agent startups is being squeezed out of existence.

The middle layer is gone

To understand the impact of this release, we first need to figure out what's new about it.

Claude Managed Agents doesn't just help you call the Claude API better; it allows Anthropic to run the Agent directly.

In the past, if developers wanted to create a practical Agent, they had to shoulder the entire infrastructure, including the agent harness, managed operating environment, built - in tools, sandbox, secure execution, and long - task operation.

But now, Claude Managed Agents takes care of it all.

What does this mean?

If a startup builds this infrastructure on its own, it might take 6 - 12 months of engineering work, plus ongoing maintenance costs.

But Anthropic now says, "Don't build it. Use mine and pay by the hour."

This is like cutting off the "middle layer" that a batch of Agent startups rely on for survival.

But Anthropic isn't the only one doing this.

Cloud providers and model vendors are collectively platformizing the Agent operating layer.

Microsoft's Foundry Agent Service has made its intentions clear. It not only supports no - code prompt agents but also allows the construction and hosting of hosted agents using Agent Framework, LangGraph, or even custom code.

In other words, Microsoft can directly host Agents.

Google is the same. The official documentation of Vertex AI Agent Engine states that the platform can handle all the infrastructure, scaling, security, monitoring, etc. of Agents, and users only need to focus on the business logic.

The signal is very clear.

The value layers that Agent startups used to boast about are being gobbled up layer by layer by upstream platforms.

Who will die first?

For startups with shallow moats that only focus on the middle layer, this will be a direct blow.

Type 1: API resellers

There are a number of domestic companies that survive by solving the problem of "unable to directly access Claude in China." They offer aggregated APIs, load balancing, and billing management.

After the release of Claude Managed Agents, the value of these services has instantly dropped to zero.

Enterprises no longer need middlemen. They only need an overseas entity to directly use Anthropic's official hosting service.

Type 2: General Agent development and orchestration platforms

This type of service is the most vulnerable to being replaced by platforms.

StackAI claims to visually drag - and - drop to build AI workflows. Its selling points are highly concentrated on "building, running, and integrating Agents." It's like a SaaS layer that wraps large - model API calls, without any real moat.

E2B sells the security sandbox and hosted execution environment for AI agents. Now that Anthropic has taken care of the most headache - inducing issues for developers, such as permission isolation, code execution, and security auditing, who would waste money on its dedicated cloud?

Dify.ai, a leading player in China's low - code Agent platform, once had a high valuation.

It features one - stop orchestration, knowledge base integration, tool marketplace, and hosted deployment. Countless entrepreneurs and SMEs have launched Agents within a few days with its help.

But now, Dify's core selling points are perfectly covered by the official service. Agents and integrators who earn enterprise service fees through Dify will soon be out of work.

These thin - middle - layer companies have made a fortune from model dividends in the past year, thinking that saving developers' engineering work was their moat.

But now, this moat has become a standard feature of upstream vendors, and such business models are about to become worthless.

Type 3: Undifferentiated orchestration frameworks

Frameworks like LangChain and CrewAI were once the default choices for enterprises to build Agents.

But now, the value of these frameworks has dropped from a necessity to an option.

For example, CrewAI started with a multi - Agent orchestration framework and cloud hosting service. It had a good run in the past year with enterprise customers paying to run complex workflows.

But now, the built - in native capabilities of Claude Managed Agents have made its simple and useful selling points into standard features.

Without stronger enterprise governance, evaluation, monitoring, industry templates, and result delivery, why would customers pay extra?

Of course, not all Agent companies will die.

Startups that focus on vertical closed - loop scenarios, such as financial risk - control Agents with their own risk - control data, medical Agents with exclusive case libraries, and industrial Agents connected to production - line IoT, can survive and even thrive.

Because they are companies with strong scenarios, strong data, and strong result delivery. Customers pay for products that are closer to business outcomes.

Where are the startup opportunities?

Over the past few years, the AI industry has been repeating the same cruel scenario.

Every time a model is upgraded, a batch of AI startups suddenly go under.

In 2023, when OpenAI released GPT - 4, a batch of startups focused on text generation died.

In 2024, when OpenAI released GPTs, a batch of startups focused on lightweight Agent platforms died.

In 2025, when OpenAI released Operator, a batch of startups focused on browser automation died.

In 2026, when Anthropic released Managed Agents, a batch of startups focused on the Agent operating layer are about to die.

It seems that a batch of AI entrepreneurs are going under, but in fact, it's those who mistake platform dividends for their own product capabilities that are failing.

This is the cruelest part of AI startups.

Many companies don't lose to their peers but to the unlit square on the upstream vendors' product roadmap.

But there are also application - layer players that can't be killed by model upgrades.

Those applications that are farthest from model capabilities and closest to customer business have industry know - how, data flywheels, in - depth workflow integration, and high customer switching costs.

The stronger the model, the more valuable these closed - loop applications become.

History repeats itself every time.

Anthropic has just sharpened the blade faster this time.

The next batch of players to be eliminated is already in line.

This article is from the WeChat official account "World Model Workshop". The author is World Model Workshop. It is published by 36Kr with permission.