TOOLS & SKILLS
Claude Managed Agents 🛠️
Anthropic just dropped something massive. You can now tell Claude what you want an AI agent to do — and Anthropic builds it, runs it, manages it, and keeps it.
What Anthropic Just Dropped
On April 8, 2026, Anthropic launched Claude Managed Agents in public beta. It's available right now through the Claude API.
Here's the short version: before this, if you wanted to build an AI agent — something that can work on its own, complete tasks, manage files, and talk to other tools — you had to build all of that yourself. The agent loop, the tool execution, the error handling, the infrastructure. That could take weeks and a lot of money.
Now, Anthropic handles all of it. You describe what you want the agent to do. They build, run, manage, and host it on their servers. If something breaks, it recovers automatically. If the task takes hours, it keeps going.
Think of it this way: before, you had to build the car, the road, and the gas station. Now Anthropic says “just tell us where you want to go.”
This is a big deal because AI agents went from something only engineers could build, to something founders with some technical skills could build, to — now — something anyone can spin up through an API call. The barrier just dropped dramatically.
The 4 Building Blocks
Managed Agents are built on four concepts. Once you understand these, everything else clicks.
Agent
The blueprint. You define what model to use, what tools it has access to, and what its instructions are. Reusable — create once, run many times.
Environment
The workspace. A secure cloud container where the agent runs code, reads files, and accesses the internet. You control what it can and can't reach.
Session
A running instance. When you give an agent a task, it starts a session. The session keeps running until the task is done — even if it takes hours.
Events
The conversation. Messages between you and the agent — you send instructions, the agent sends back results, status updates, and questions.
What Your Agent Can Do Out of the Box
Every Managed Agent comes with 8 built-in tools enabled by default. No setup required.
- Bash — Run shell commands (install packages, process data, run scripts)
- Read — Read any file in the workspace
- Write — Create and write files
- Edit — Make precise edits to existing files
- Glob — Find files by pattern (e.g., all .csv files in a folder)
- Grep — Search inside files with regex
- Web Search — Search the internet for real-time information
- Web Fetch — Pull content from any URL
On top of these, you can add custom tools (your agent calls your code when it needs something specific) and MCP servers (connect to external services like databases, APIs, or internal tools using the Model Context Protocol).
How to Create Your First Managed Agent
You need a Claude API account. If you've never used the API before: go to console.anthropic.com, create an account, add a payment method, and grab your API key. That's it.
There are two ways to do this: the CLI tool (fastest) or API calls (more control). I'll show you both.
Create an Agent. This is the blueprint — what model to use, what instructions to follow, and what tools it has access to. You define its name, system prompt (its personality and rules), and which Claude model it runs on. Available models: Claude Sonnet 4.6, Claude Opus 4.6, and anything Claude 4.5 or later.
Create an Environment. This is the secure container where the agent works. You can pre-install packages (Python libraries, npm packages, etc.), control network access (restrict which websites it can reach), and mount files it needs. Think of it as setting up a desk for your new employee.
Start a Session. Connect your agent to its environment and give it a task. The session starts, the agent begins working, and you receive real-time updates as events. The agent keeps going until the task is done or it needs your input.
Stream the results. You can watch what the agent is doing in real time through Server-Sent Events (SSE). You'll see its thinking, its tool calls, its results — everything. When it's done, it goes idle and waits for your next instruction.
What People Are Actually Building
Data Processing
Upload a messy CSV, the agent cleans it, analyzes trends, builds charts, and delivers a summary report — all autonomously.
Research Agents
Give it a topic, it searches the web, reads sources, cross-references data, and writes a comprehensive brief with citations.
Code Generation
Describe a tool you need, the agent writes it, tests it, debugs it, and delivers working code — complete with documentation.
Content Pipelines
Feed it raw notes, it produces blog posts, social captions, email newsletters, and repurposed formats — in your voice.
Multi-Agent Teams
One agent coordinates others. A manager agent delegates research to one agent, writing to another, and review to a third.
Internal Tools
Connect to your company's APIs via MCP, let the agent query databases, update records, and generate reports on a schedule.
What It Costs
Two costs: tokens (what Claude thinks) and session time (how long the container runs).
Managed Agents vs. Other Options
Where does this fit in the Claude ecosystem? Here's the honest breakdown.
Why This Matters Right Now
AI agents started showing up less than a year ago. At first, only engineers could build them. Then it was technically-savvy founders. Now, Anthropic is saying: we'll handle the hard parts — you just tell us what the agent should do.
Every month, the barrier to entry gets lower. Every month, what agents can do gets more powerful. The people who start learning how agents work now — even at a basic level — will have an enormous advantage in 6 months when this technology is everywhere.
You don't need to be an engineer. You don't need to understand containers or API architecture. You need to understand what's possible and how to think about automation. That's the skill that transfers no matter which tool you use.
The gap between “person who uses AI” and “person who has AI working for them” is closing fast. Managed Agents just closed it further.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.