Anthropic's new product is so powerful that it could make the AI Agent Infrastructure team obsolete?

Original Article Title: "Anthropic Launches a New Product Today, Potentially Rendering a Wave of AI Agent Infrastructure Teams Obsolete"

Original Article Author: Bayun, AI Researcher

This product is called Claude Managed Agents. In a nutshell: you tell Anthropic what kind of AI agent you want, and they help you run it in the cloud, providing full infrastructure support with usage-based billing. Sentry used it to automate bug fixes in a few weeks, while Rakuten deployed a specialized agent in one week. Previously, these tasks would require an entire engineering team working for months.

At the same time, Anthropic's annual recurring revenue has just surpassed $30 billion, tripling from last December. Most of the growth comes from enterprise customers. Wall Street is already getting nervous, with the WSJ reporting that investors are becoming increasingly cautious about the stock prices of traditional SaaS companies, fearing that products like Anthropic's could make some traditional software services redundant.

What exactly is this product? How does it differ from the Claude Code you may already be using? And how was it achieved technically?

What Is It? How Does It Differ from Claude Code?

If you have used Claude Code, you know how AI agents operate: you give them a task, and they autonomously plan steps, use tools, write code, modify files, and complete the task step by step.

Claude Code runs on your own computer and is a personal command-line tool for developers. It stops running when you shut down your computer.

Managed Agents run on Anthropic's cloud and are an API service for enterprises. They can run continuously 24/7, retain progress even if disconnected, and your product can directly embed AI agent capabilities.

This is how Notion operates: users assign tasks to Claude agents within Notion, the agents work in the background, complete the tasks, and return the results without users ever having to leave Notion.

Several Typical Use Cases:

· Event-Triggered: The system discovers a bug, automatically assigns a bot to fix it and raise a PR, with no human intervention in between.

· Time-Based: Automatically generate a GitHub activity summary or team work briefing every morning.

· Fire-and-Forget: Assign a task to a bot in Slack, it completes the task and returns the document, slides, or app.

· Long-Running Task: Running in-depth research or code refactoring for several hours.

What Makes Cloud-Native Agents Different from In-House Ones?

You can build your own, but it's costly and slow.

An AI agent that can go live requires much more than just "making an API call": a sandbox environment (an isolated secure space where AI can run code, modify files without affecting the real system outside, like providing a dedicated virtual computer to AI), credential management, state recovery, access control, end-to-end tracing, and more.

Previously, many enterprise customers needed an entire team of engineers dedicated to these tasks. Now, it's plug and play, allowing engineers to focus on the core of the product.

However, the pain points addressed by Managed Agents go beyond just saving manpower.

Mason Dongxi (@dongxi_nlp) has a concise summary:

There is a specific example in the Anthropic Engineering Blog:

When Claude Sonnet 4.5 approaches the limit of the context window, it starts to "worry" and hastily ends the task. They added context reset in the scheduling framework to address this. However, with the release of Claude Opus 4.5, this issue disappeared, and the previous patch became a burden.

When you build your own scheduling framework, you have to update it with every model upgrade. Entrust it to Anthropic, they optimize it for you, technically optimizing what they sell to you.

Who's Using It? How?

Notion allows users to delegate tasks such as coding, creating PPTs, and organizing spreadsheets directly within the workspace to Claude, running dozens of tasks in parallel, with the entire team collaborating on the same output. Notion Product Manager Eric Liu said users can delegate open-ended complex tasks directly without leaving Notion.

Sentry implemented a fully automated process "from discovering a bug to submitting fix code." After their AI debugging tool Seer identifies the root cause of an issue, Claude directly writes patches and opens PRs (pull requests). Engineering Director Indragie Karunaratne said they were able to go live in a few weeks, saving the operational cost of maintaining custom infrastructure.

Atlassian integrated it into Jira, allowing developers to directly assign tasks to the Claude AI agent.

Asana created AI Teammates, integrating AI collaborators into project management to receive tasks and deliverables.

General Legal (legal tech company) implemented a unique approach: their AI agent can temporarily develop tools to search data based on user queries. Previously, each user query required anticipating and developing a search tool in advance, but now the AI agent generates them on-demand. The CTO said development time has been reduced by 10 times.

Rakuten deployed dedicated AI agents in various departments including engineering, product, sales, marketing, and finance, each going live within a week, receiving tasks through Slack and Teams and returning deliverables such as spreadsheets, PPTs, and apps.

Technical Principle: Decoupling the Brain from the Hands

The Anthropic engineering team wrote a technical blog post titled Scaling Managed Agents: Decoupling the brain from the hands, discussing the architecture evolution behind Managed Agents.

Initially, they crammed everything into one container: AI's inference loop, code execution environment, and session recording, all together. The advantage was simplicity, but the disadvantage was that all the eggs were in one basket. If the container crashed, the entire session was lost, and there was no way to replace a specific part individually.

Later, they made a key split:

· The "Brain" was Claude and its scheduling framework, responsible for thinking and decision-making.

· The "Hand" was the sandbox and various tools, responsible for executing specific operations.

· The "Memory" was an independent session log, recording everything that happened.

The three were independent of each other, and if one crashed, it did not affect the other two.

This split brought several practical benefits:

Speed

Not every task needs to start the full sandbox environment. Now, the sandbox is only started on-demand when AI actually needs to run code. The median first response latency decreased by about 60%, and in extreme cases, it decreased by over 90%.

Security

Code generated by AI runs in the sandbox, while credentials to access external systems are stored in a secure vault outside the sandbox, with physical isolation on both sides. For example, when accessing a Git repository, the system initializes by cloning the code, and AI normally uses git push/pull, but the Token itself is not visible to AI. For services such as Slack and Jira, they are accessed through the MCP protocol, requests go through a proxy layer, the proxy layer retrieves credentials from the vault to call the service, and AI never handles the credentials throughout.

Flexibility

The Brain does not care what the Hand is. There is an interesting sentence in the engineering blog: the scheduling framework doesn't know if the sandbox is a container, a mobile phone, or a Pokémon emulator. As long as it fits the interface of "input a name, get a string out," it works.

This also means that multiple Brains can share the same Hand, one Brain can hand over the Hand to another Brain, laying the foundation for multi-agent collaboration.

Limitations

Managed Agents are not omnipotent. There are several points to note:

Some features are still in the research preview stage. Abilities such as multi-agent collaboration, advanced memory tools, and self-assessment iteration (allowing the agent to judge the quality of task completion and iteratively improve) are not fully open yet and require application for access.

Platform Lock-in. Opting for Managed Agents means your agent infrastructure is tied to the Anthropic ecosystem. If you want to switch models or platforms in the future, the migration cost is non-negligible.

Context management remains a challenge. Although session logs are stored independently, decisions on what information to retain or discard during long tasks still involve irreversible choices. This is an ongoing challenge, and their current approach separates context storage from context management: storage ensures no loss, while management policies adjust with model evolution.

Cost Predictability. $0.08 per session hour may not sound like much, but for complex tasks that run for several hours, considering token consumption and runtime costs, the total expenses could be significant. Enterprises need to conduct budget assessments.

Managed Agents indicate that most enterprises still have a long way to go before "fully leveraging AI agents for work."

While the infrastructure barrier has been lowered, Managed Agents cannot help with defining good tasks, designing effective workflows, or establishing trust to enable AI access to core business data.

The "AWS Moment" of AI Agent Infrastructure

Managed Agents seem to be following the path AWS took years ago: first focusing on computing power, then wrapping the runtime environment around it.

A decade ago, companies debated whether to "move to the cloud," and now they are debating "self-built Agent infrastructure versus managed hosting." Historical experience tells us that most companies ultimately choose hosting because infrastructure has never been a core competency. OpenAI has also launched its own Agent platform, Frontier, and the competition in this space is just beginning.

From a technological perspective, the "separation of brain and hand" architecture is worth noting. It allows each part of the system to evolve independently: upgrade the model, swap out the brain; need a new tool, add a hand; change the storage solution, replace the memory layer.

The analogy in the engineering blog is spot on: the operating system's read() command doesn't care whether it's dealing with a 1970s disk or a modern SSD; the abstraction layer remains stable, allowing the underlying implementation to be swapped out easily.

From a user perspective, if you are an enterprise developer looking to embed AI agent capabilities into your product, Managed Agents may save you months of infrastructure work.

Six languages (Python, TypeScript, Java, Go, Ruby, PHP) are supported by SDKs. If you are already using Claude Code, update to the latest version, and run /claude-api managed-agents-onboarding to get started.

For regular AI enthusiasts, the most immediate impact might be: in the SaaS products you use, more and more AI agents will be working behind the scenes to assist you, with many of these agents likely running on Managed Agents.

Pricing Reference: Token costs follow the Anthropic API standard pricing, with a runtime cost of $0.08 per session hour (idle sessions are not billed), and $10 per thousand web searches.

Do you think that AI agent infrastructure will eventually be dominated by a few major players, similar to how cloud computing is today?

Original Article Link