AGENTIC AI

Building something that didn't exist yet

The Agentic AI project at Oracle was a zero-to-one effort. There was no existing pattern to extend and no prior art inside the product to reference. The starting point was a clear user goal:

As a developer, I need to be able to leverage my integrations to automate business processes.

That statement sounds simple, but it carries significant complexity. Developers on Oracle Integration Cloud had already built connections to dozens of external systems. The challenge was giving them a way to expose those connections to an AI agent in a form the agent could actually act on, with the right inputs, the right outputs, and the right guardrails built in.

Seeing what the agent is actually doing

Once a tool is configured and wired to an agent, the next critical question is: how do you know it's working? Testing agentic systems is fundamentally different from testing a form or an API call. The agent reasons through a sequence of steps, and any one of those steps can go in an unexpected direction. Making that reasoning visible and actionable was a major design challenge.

We went through several rounds of design exploration before landing on the right approach. Early versions tried to surface too much at once, burying the signal in noise. Later iterations swung too far toward simplicity and lost the detail developers needed when something went wrong. The best progress came from design reviews that used the artifacts themselves as the starting point for conversation, bringing engineers, PMs, and designers into the same discussion with specific screens to react to, rather than abstract descriptions of what the interface should do.