Architecture & Extensibility

Build It Yourself or Rent It? Self-Hosted Runtimes vs. Managed Agent Platforms

Self-hosted runtimeYou build, operate andscale it — full controlManaged platformA provider hosts it — lesscontrol, more leverage

A declarative markdown runtime and a Python state-machine runtime are both harnesses your team builds and operates directly. A managed agent platform is the third point on that spectrum: an outside provider hosts the runtime instead, and the trade you're making shifts from "how much control do I want" to "how much operational burden am I willing to hand off."

What actually changes

A managed agent platform — Google's Vertex AI Agent Engine is a concrete example — takes over three concerns a self-hosted runtime otherwise leaves entirely to your own team. The first is development and packaging: a self-hosted setup requires writing and containerizing the runtime yourself, while a managed platform is typically framework-agnostic, letting you bring agent code written in LangGraph, Google's Agent Development Kit, CrewAI, or similar, largely as-is. The second is infrastructure operations: instead of your team provisioning, scaling, patching, and securing the compute the agent runs on, the provider auto-scales and secures the hosting layer on your behalf. The third is observability and evaluations: rather than wiring up tracing, logging, and eval harnesses by hand, a managed platform typically ships built-in tracing, session state, and eval tooling out of the box.

What doesn't change

It's worth being precise about what a managed platform does not take off your plate: you still own the agent's own logic — what it decides, how it reasons, what tools it calls and why. The platform owns the operational burden that surrounds that logic, not the logic itself. Moving to a managed platform changes the harness, not the agent.

Making the trade-off concrete

The honest way to frame this choice is not "which is better" but "how much of our own engineering time do we want going into agent infrastructure, as opposed to agent behavior." A team early in building a new capability, without dedicated platform engineers to spare, often gets more value shipping on a managed platform and spending its scarce engineering time on the agent's actual reasoning and tool design. A team with existing infrastructure investment, specific compliance or data-residency requirements a general-purpose managed platform doesn't accommodate, or a genuine need for low-level control over the execution environment, is often better served self-hosting, accepting the operational cost in exchange for that control.

As with the choice between deterministic and non-deterministic invocation, there's no universally correct answer here — only a decision that should be made deliberately, with the actual trade-off named, rather than by default.

Questions to bring to your team

  • How much engineering time is currently going into agent infrastructure rather than agent logic?
  • Would you trade some control for a platform that handles scaling and tracing for you?

Also available in Čeština

← What Is a "Harness"? The Hidden Layer Behind Every Agent Keeping Humans in the Loop: Where Oversight Still Matters →