What is a managed agent, and do you need one?
On this page
The Red Laptop
Growing up, my dad ran his business off a database on a red laptop in his office. I wasn't allowed to touch it. When I told him to use a managed database like RDS instead, he didn't want to hear it. The laptop worked, and it was cheap.
Then came the upkeep: keeping the disk from filling up, keeping the laptop on, keeping the home internet running. One day he had to wipe it to make room, and his company was down for an afternoon. A few months later, he moved to AWS.
The laptop wasn't a mistake at first. The cost just showed up later, in work that had nothing to do with his business. Almost every team building AI agents today has its own version of that laptop.
You built the agent. Now where does it run?
If you're building an agent today, you probably started with something like LangChain, Mastra or the Claude Agent SDK. They're great at helping you build the agent. You give it a prompt and some tools, and it works on your laptop.
But on their own, they don't run it for you. A team I talked to recently put it almost exactly like this: "we built it locally, but we don't know how to deploy it."
Here's what usually happens next.
- You put it on a server. Now people need a way to reach it, so you build an API.
- The server can't run all day for an agent that's mostly waiting. So you shut it down when it's idle, bring it back when someone replies, and make sure it remembers where it was.
- Things break. A server restarts halfway through a task. A model call times out. The agent has to reliably pick up where it left off.
- Real users show up. Each person's conversations have to stay separate and reach the right place. Each customer has their own logins, and none of them can sit where the agent could read them.
- You need to see what's going on. Which agents are running, what they did, what they cost, and how to stop one customer from burning through your budget.
And it goes on from there. None of this is impossible. There's just a lot of it, each piece affects the others, and none of it is what your customers pay you for. It's my dad's laptop all over again.
Four ways to build an agent
There are four main ways to build an agent today:
Managed agent platforms are the managed database of this picture: you still build the agent, you just stop babysitting where it runs.
This is a real category now. Anthropic launched Claude Managed Agents in April, and OpenAI launched its Agents API in September.
We wrote a longer guide to the four options if you want the details.
Do you need a managed agent platform?
Not always. If your agent runs a mostly fixed process, a workflow builder is fine. If it's one agent doing one job on one machine, a framework or harness may be all you need.
You need a managed agent platform once your agent serves many users, runs for a long time, or waits on people. That's when the five steps above start to bite.
Which platform should you pick?
Honestly, it depends.
- If you only want Claude models and you're happy on Anthropic's cloud, Claude Managed Agents is a good choice.
- Same with OpenAI's Agents API if you're all in on OpenAI.
- If you want to switch models as better ones come out (including open-source ones), choose where your agent runs, keep your data on your own servers, or spend less, that's why we built Omnara. It's open source, so you can run it yourself or let us run it.
We keep a longer comparison of the platforms, including ours.
The end of the laptop
Nobody starts a company on a laptop database today, because someone else made running databases boring. We want to make running agents just as boring.
If you're somewhere in the middle of this, we're happy to chat.