Skip to main content

Custom production AI agents

Some AI applications outgrow configuration: they need custom logic, specific dependencies, their own data stores, and control over how they scale. This page explains when an AI feature warrants a purpose-built agent, how DIAL Custom Apps let you build one with full control while staying inside the platform, and what that costs. It is for architects sizing the approach and for the developers who would build it.

Why this matters​

A no-code tool is the right answer until it is not. The moment an application needs a library that is not on the approved list, a connection to a database you own, business logic too intricate for a prompt, or a scaling profile you control, configuration-based building runs out. The honest options are then to build a standalone service outside any platform, or to build inside the platform and inherit its governance.

Building outside means re-implementing what the platform already does: identity, access control, cost limits, model access, observability. Each of those is a project on its own, and skipping any of them is how an ungoverned service ends up in production. The value of a platform-native custom agent is keeping full control over your code while not rebuilding the surrounding plumbing.

What you can do with DIAL​

A Custom App is a containerized application you write in Python with the DIAL SDK. You own the container, its dependencies, and its network access, and you deploy it on your own infrastructure. It speaks the Unified API like any other application, so it reaches models through DIAL Core and is governed the same way a Quick App is.

Because it is a full application, it can do what configuration cannot: call your databases, integrate external systems, run arbitrary logic, and act as an agent that calls other models and applications through the Unified API to coordinate multi-step work. It is also reusable: a Custom App can be the business-logic tool that a no-code Quick App orchestrates.

How you'd build it​

Why DIAL over the alternatives​

A framework such as LangChain gives you primitives to build an agent, but it does not give you a place to run it under governance — you still choose where it lives, how it authenticates, and how its cost is tracked. A standalone service gives total freedom and total responsibility. A DIAL Custom App keeps the freedom that matters — your code, your dependencies, your scaling — while the platform supplies identity, access control, model access, and observability. The trade-off is that you self-manage the container rather than letting the platform host it, which is the price of full control. When to use which app type compares this against the managed options.

Further reading​