What Are Managed Agents in AI Studio?
Managed agents in AI Studio are AI helpers that Google runs for you in the cloud. You type one request, and Google creates a full agent that can think, plan, search the web, build files, and finish a task from start to end. You don't need a server or technical setup—just ask.
The agent gets its own computer with Ubuntu Linux, four CPU cores, and 16 gigabytes of memory. The brain is Gemini 3.5 Flash, the same engine Google uses for its own agents. Everything runs on Google's side, so your laptop never gets touched.
These agents are now in public preview, which means anyone can try them right now. This is a big change from before, when building an AI agent required servers, connecting tools, and technical glue. Now it's just one request.
Main Features of Managed Agents
Managed agents come with several powerful features that make them useful for real work. Here are the main ones from the episode:
- One-request setup: You type a request, and the agent spins up with its own cloud computer.
- Plain-English instructions: You define the agent with two text files—agents.md for instructions and skill.md for skills—no code needed.
- Real file output: The agent can build files like guides, summaries, or sign-up pages, and you can download them.
- Memory with session IDs: The agent's cloud computer gets an ID, so you can reuse it. It sleeps after 15 minutes of inactivity and stays for a week.
- Safety controls: You can restrict the agent's web access to a list of allowed sites, and passwords are handled outside the agent's environment.
- Save and reuse: Once you set up an agent, you can freeze it and call it by name anytime.
Google also stores the agent's files and installed tools, so you can pick up where you left off.
How to Set Up Your First Agent
Setting up a managed agent in AI Studio is straightforward. Here are the steps shown in the episode:
- Go to Google AI Studio.
- Pick the managed agent option.
- Type your request in plain words—for example, "Find the most common questions new business owners ask about AI automation and turn them into a short welcome guide."
- Let the agent run. You can watch it think and work in real time.
- Create your agents.md file with instructions like "You are my research helper. Always keep things simple. Always save your work as a clean file."
- Create your skill.md file to define a skill, like "When I ask for slides, build me a simple slide deck."
- Run the agent. It runs on Google's cloud, and you get the result.
If you're more technical, you can also run it through code with a short command that specifies the agent, your request, and a remote sandbox. The agent remembers your session, so files and settings stay intact.
Real Tasks You Can Hand to an Agent
In the episode, Julian showed several real tasks a business owner would give a managed agent. These are simple, everyday jobs that can save hours:
- Research questions: Ask the agent to find common questions about AI automation and write them into a welcome guide.
- Weekly news summary: Tell it to search the web, read top stories, and write a one-page summary in plain words.
- Build a sign-up page: Request a clean one-page sign-up that explains value and encourages joining.
These examples show you don't need to code—you just describe what you need. The agent goes off, does the work, and hands you a real file you can download.
This is the key difference from chat responses: agents finish the job. They don't just talk; they produce a usable result.
Results and Comparisons in the Episode
The episode highlighted how fast managed agents have evolved. Google's first managed agent, Deep Research, came out in December 2025 and could only write reports with hundreds of searches. In April, Google launched a bigger platform for companies, but it was locked down. Then on May 19, 2026, at Google I/O, they opened it to everyone inside AI Studio.
Google's CEO Sundar Pichai said Google now handles 3.2 quadrillion AI tokens every month, up from 480 trillion a year earlier—a nearly seven-fold jump. This shows how widely AI is being used.
One early-access developer said going from a prompt to a full agent in one call felt "almost magical." His team set up two real agents the same day they got access. Philip Schmid also explained how the hard work moved to Google, letting users focus on what the agent should do, not how to run it.
Tips, Limits, and Mistakes to Avoid
Managed agents are new and still in preview, so they're not perfect. Julian shared honest advice:
- Check their work: Agents make mistakes and can go off track. Always review what they produce.
- Write clear instructions: A lazy instruction gets a lazy result. Be specific about what you want.
- Use safety features: Set allowed websites to keep the agent on a leash, and keep passwords out of its reach.
- Start with small tasks: Give the agent one real task from your week to see how it works before scaling up.
Once an agent is set up, you can save it and reuse it. It starts clean each time but keeps your instructions, so you build once and use forever.
Because it's in preview, expect rough edges. But the people who learn early will get a head start over those who wait.
Who Managed Agents Are For
Managed agents are for anyone—not just coders. The whole point is to remove the technical barrier. Business owners, marketers, and curious individuals can all use them because you interact in plain English.
Julian stresses that the skill that matters now is knowing what to ask for, not coding. You point an agent at a real problem in your business and get back something useful.
This is especially valuable for small business owners who are busy. You can hand off research, summaries, and page building, which frees up hours every week for customers, selling, and other high-value work.
The setup excuse is gone—Google removed it. Now the only thing holding you back is learning how to use them, and that's something you can control.
What's Next for Managed Agents
Google has already promised more features. They're building ready-made templates inside AI Studio, so you won't have to write instructions from scratch. Instead, you'll pick a starting point—like a research agent or a page builder—and tweak it.
This will make it even easier. Right now you write a few plain-English lines; soon you might just pick from a menu. Agents will also get faster brains, more tools, and more memory.
The trend is clear: from one locked-down agent in December to wide-open agents by May. That curve isn't slowing down.
Julian's final advice: try one agent today, give it one small real task, and watch how it works. Then learn to write good instructions and point agents at tasks that eat your time. Early adopters will move way faster than everyone else.
