Thinking about building an AI agent but not sure where to start? Oracle AI Agent Studio makes it easier than ever to create custom agents that can handle real work inside Oracle Fusion Cloud Applications—from triaging service requests to pulling real-time data from your ERP. But like any good tool, getting the most out of it takes a little planning. This guide walks you through the basics: how to scope the right kind of problem, how to design an effective team of agents, and what tools to give them so they can get the job done.

Focus on the problem

Building an agent is similar to establishing a new team of people who will work together to perform a particular business function. Nobody would fund you to hire a team of five people if you didn’t define what problem they were going to solve, who they are solving it for, and what degree of accuracy is needed for the solution. The same is true of agent teams.

So stay tightly focused on the job. And remember, agents are not the best solution to every problem. Make sure the challenge you’re tackling is one that leverages AI agents’ particular strengths (i.e., working with ambiguity, decision making, coordination of subtasks, etc.). As you start to map out what your agent team will do, keep on thinking back to whether the problem and the instructions you provide would also be understandable and achievable by a human team member. If not, then you’ll need to provide more tools or more data—or rethink the task.

Design your team

Just as with real team members, agents become more efficient when they’re allowed to specialize. And dividing up tasks helps to ensure quality standards are met. If you give a single agent too much to do, you end up with no way to optimize all the different moving parts needed to get a good result.

Best practice is to divide tasks between different worker agents, which take instructions from a supervisor agent. The supervisor agent reviews tasks and delegates them to the worker agents, then reviews output from the worker agents and further assigns additional work as needed until the job is complete. The supervisor agent can even make decisions if conflicts arise between worker agents.

Let’s look at a simple example. Let’s say we’re building an agent team to respond to service requests. We could use a:

  • Worker agent that checks for resolutions in a knowledge base
  • Worker agent that checks previous ticket resolutions
  • Supervisor agent that compares their outputs, determines which one (or what combination of both) is going to best help the user, and writes a response

If the user comes back and says the resolution didn’t work, then the supervisor agent can layer additional information from the user onto the request and ask the workers for another pass.

Tools help workers complete tasks

Large language models power agents and are good at many things—but not everything (e.g., math). Fortunately, agents can be given tools to use when they are asked to do something they aren’t good at. Examples are:

  • Document tool: enables the agent to refer to a PDF or Word doc containing authoritative info relevant to a question (i.e., retrieval-augmented generation, or RAG)
  • Business object tool: enables the agent to query (or write to) tables in the Fusion Database to pull information relevant to a given person or entity (Oracle’s architecture ensures that the agent only gets the same access as the user would have ordinarily)
  • Third-party APIs: enables the agent to call out to other APIs (e.g., for currency exchange rates, weather conditions, or search engine results)
  • Calculator: helps the agent do math
  • Messaging tool: lets the agent send emails (inbound email and more messaging apps coming soon)
  • Deep link tool: enables the agent to generate links to other objects within Fusion Apps

Consider your sources of truth

Hallucinations occur when an agent is asked to output more information than it has real knowledge about. It makes up content to fill the gaps. Clearly, you want to avoid this as much as possible. Before starting to build anything, it can be very helpful to map out what the sources of truth will be for the factual information the agents will base their output on.

One of the best sources of truth is always going to be a company’s official policies or procedures. A retrieval augmented generation (RAG) step can allow an agent to review official sources of information and ensure its responses come only from that source of truth. As mentioned above, the web search tool provides the option to get information online, but it should be used with caution given the uneven quality of information from internet sources.

When it comes to business-critical information, it’s far better to give the agent a source of truth that a business already trusts. In many cases, that’s going to be the Fusion Apps database the agent already sits on. That’s where the Business Objects Tool comes in. It allows you to quickly and easily give the agent access to the same information in the Fusion database as the user of the agent would ordinarily have. It’s great for tasks like looking up leave/absence balances, or finding the latest statuses of shipments.

Conclusion

AI Agent Studio makes it easy to set up a team of agents to do work. But, just like with human teams, you won’t get the result you want if the agent team isn’t organized smartly and around a goal its likely to be successful at. Following best practices can help you get the most from your team.

Additional resources

Related posts you might like

 

Get more information link

If you’re an Oracle customer and want to get new stories from The Fusion Insider by email, sign up for Oracle Cloud Customer Connect. If you’re an Oracle Partner and want to learn more, visit the Oracle Partner Community.