Managed Agents vs Self-Hosted Agent Loops: How to Choose
In 2026 you can rent the entire agent runtime — the loop, the container it acts in, scheduling, even a grader that keeps the agent working until the output meets your rubric. Or you can host all of it yourself. The choice is less about capability than about who you want holding the keys.
- Published
- Updated
Side by side
| Dimension | Managed agents | Self-hosted (SDK on your infra) |
|---|---|---|
| Loop and sandbox | Vendor runs both | You run both |
| Time to first working agent | Days | Weeks |
| Scheduling / cron | Built in | You build it |
| Persisted, versioned agent configs | Built in | Your repo and CI |
| Outcome grading against a rubric | Built in | You build the grader |
| Data residency | Vendor region options | Anywhere you choose |
| Custom runtime / private network | Limited | Unlimited |
| Cost model | Session-based plus tokens | Tokens plus your compute and engineers |
Choose managed when
- You want an agent that runs on a schedule — nightly reports, weekly reconciliations — without maintaining a scheduler.
- The task is long-running and file-heavy, and you would rather not operate containers.
- You need "work until it is right" behaviour and want a grader you did not have to write.
- Your team is small and speed to production matters more than runtime control.
Choose self-hosted when
- The agent must reach systems inside a private network with no public route.
- Compliance dictates where data and compute live, down to the rack.
- You need a custom tool runtime — GPUs, proprietary binaries, unusual languages.
- You already run agent infrastructure and the marginal agent is cheap for you.
A common middle path: managed for the loop, with sensitive actions implemented as custom tools whose execution stays on your side. Credentials never enter the sandbox; the managed runtime only sees results.
What we recommend
For most businesses shipping their first serious agent: managed. You will learn what the agent needs to do in weeks instead of months, and the config is portable enough to move later. We build on both, migrate between them, and teach teams the trade-offs in a half-day session before they commit. Describe the agent and its data constraints below and we will tell you which side we would put it on.
Frequently asked questions
Is a managed agent locked in?
The agent definition — system prompt, tools, model — is portable. The scheduling, sandbox and grading features are what you would need to rebuild if you moved.
Can a managed agent call my internal APIs?
Yes, through custom tools where the execution happens on your side, or through MCP servers you expose. Secrets can be held in a vault and substituted at the edge rather than placed in the sandbox.
Which is cheaper?
For a handful of agents, managed — the engineering time saved dominates. At large scale with existing infrastructure, self-hosted can win on unit cost.
