In partnership with
What's your current approach to AI agents?
Building an agent with the Claude API used to mean building the whole thing yourself.
So you’d have to set up the agent loop, tool execution layer, sandbox, file system, and the session state.
Anthropic shipped something last week that skips all of it. It's called Claude Managed Agents
It gives you a fully managed runtime where Claude can run bash commands, read and write files, browse the web, and execute code inside a secure container.
You define the agent and the task, the infrastructure is taken care of for you.
It's still in beta but I gave it a shot and it looks promising. Before we dive in, something worth checking out:
Tools of the Week
1. Cursor 3
Cursor rebuilt its entire interface around a new Agents Window that runs multiple agents in parallel across local, cloud, and remote SSH environments simultaneously. The IDE view still exists, but their bet is on the developer's job to orchestrate agents, not write every line.
2. Claudoscope
Open-source tool that tracks costs, sessions, and usage across all your Claude Code runs with a searchable dashboard. If you're going to start paying $0.08 per session hour, this is how you avoid a surprise at the end of the month.
Someone reverse-engineered exactly what happens when you type into Claude Code and mapped the full agent loop, 50+ tools, multi-agent orchestration, and unreleased features from source.
What is Managed Agents?
Instead of building your own agent loop, you get a managed harness that handles execution for you.
Four things that exist in this system:
Agent: the model, system prompt, and tools. Create it once, reference it by ID every time.
Environment: a cloud container that runs on Anthropic's infrastructure. Bash, file read/write, web search, and web fetch are all built in.
Session: one running instance of your agent doing one task.
Events: everything streaming between your app and the agent in real time.
The container is the key part.
Claude can write a Python script, execute it, read the output, and iterate inside the same session, on Anthropic's servers, without you touching a single line of infrastructure code.
Building a Real Agent
Here's a concrete example: an agent that takes a company name, searches for recent news and competitor activity, writes a structured brief, and saves it as a markdown file. The kind of thing you'd do manually before a sales call.
Prerequisites: Python 3.8+, Anthropic Python SDK, API key.
Create a file called research_agent.py and paste this in:
This is a preview image, copy the entire code below 👇
Run it:
What you see in the terminal:
Output:
Claude searches the web, fetches pages, writes the file, done.
The agent and environment are reusable i.e. pass a different company name and a new session starts on the same infrastructure. You're not recreating anything.
What It's Not For
If you need tight control over every turn of the conversation, or you're intercepting tool calls for your own logic, the Messages API is still the right choice.
Managed Agents is for tasks that run long, touch the real world (files, bash, the web), and don't need you watching them the whole time.
It's in beta, so behaviors can change between releases. Multi-agent coordination, outcomes, and memory are still in research preview and you have to request separate access.
Don't build something production-critical on this today without accounting for that.
My Take
The thing I keep coming back to is session persistence. State is stored server-side.
Anthropic basically said: the reason most agents never ship is that developers spend months on infrastructure before writing a single line of product logic. So we'll just take that off your plate entirely.
That's the right diagnosis and for most teams, the right fix.
The $0.08/hour sounds cheap until you run a research agent for 8 hours and realize the runtime cost just overtook your token cost.
Nobody is talking about that tradeoff right now because everyone's excited about the demo, but worth doing the math before you build something that runs continuously.
The beta caveats also matter more than Anthropic is making them sound.
Multi-agent coordination is still a research preview. What you can access today is essentially a well-managed single agent, which is useful but not the full vision they're selling.
My honest read: if you've started building an agent and abandoned it at the "secure sandbox" step, this removes the exact wall that stopped you and you should try it this week.
If you're evaluating this for something running at scale, wait for the beta label to come off.
Until next time,
Vaibhav 🤝🏻
If you read till here, you might find this interesting
#Partner 1
TPC and Bethpage Use This Tech to Modernize Golf Courses
In golf, names like TPC and Bethpage evoke greatness. In business, they’re just a few names using Edison Interactive to revolutionize guest experiences. Edison combines next-gen content like live sports, streaming, and more with AI-powered personalization to maximize engagement. It’s how Edison earned $60M in revenue. Now, partners like Verizon are helping Edison reach high-growth verticals like pro sports and beyond.
This is a paid advertisement for Edison Interactive Regulation CF offering. Please read the offering circular at https://invest.edisoninteractive.com/
#Partner 2
LLM traffic converts 3× better than Google search
58% of buyers now start their research in ChatGPT or Gemini, not Google. Most startups aren't showing up there yet.
The ones that are get cited by the AI tools their buyers, investors, and future hires already use. And they convert at 3×.
Download the free AEO Playbook for Startups from HubSpot and get the exact steps to start showing up. Five minutes to read.
