How to Create an AI Agent in 7 Steps

Businesses are working to get more value from generative AI by incorporating it into everyday business processes. They’re starting to deploy software called AI agents in a range of applications and they’re designed to hold written and spoken conversations and to query databases to execute multistep tasks—without being programmed in advance about how to proceed in all situations.

Agentic artificial intelligence applications can be deployed to work step by step to gather the information they need by interacting with computer users and their calendars, tapping information in local and cloud systems, and using search engines or other websites to answer questions or execute actions. They rely on underlying AI large language models (LLMs) for their predictive power and ability to interact with human users in natural language.

Providers of productivity, customer management, and back-office applications have begun furnishing their customers with design studios for customizing, instructing, and activating AI agents—or creating their own. Here are a set of useful guidelines on how to outfit an AI agent for duty, including seven steps to getting an agent built and in the field.

What Are AI Agents?

AI agent software harnesses LLMs trained on vast amounts of data to find relationships and make associations among concepts, which can make relevant predictions about computer users’ intended meaning, communicating in conversational language. Agents are designed to mediate between users and language models, actively taking steps to address problems in a multitude of areas.

They can help organizations automate repetitive processes, such as assisting with financial forecasts, helping HR teams navigate multiple steps in a recruiting process, or summarizing account information and identifying upsell opportunities for sales reps.

How to Create an AI Agent in 7 Steps

AI agents are intended to understand users’ organizational roles, draw on data from business documents so that workflows remain relevant, and respond to natural language prompts instead of precoded instructions. To prepare them for that flexibility under changing circumstances, organizations need to do some prep work.

1. Choose your agent-building strategy. Businesses need to decide up front whether they want to customize prebuilt agents supplied by software vendors to help automate processes or build their own agents from scratch. Given the early stage of industries’ AI agent testing and rollout, most businesses will likely customize prebuilt agents to be better positioned to begin realizing value. When making the decision, organizations should consider the following:

• AI talent on staff: Designing custom agents requires AI developers, data scientists, and user interface experts to do the necessary programming and systems integration, whereas application administrators can work in a design studio environment to customize off-the-shelf agents.
• Model training expertise: Most businesses won’t have the in-house knowledge to select an LLM, necessary when developing AI agents from scratch, and to conduct the fine-tuning needed to make sure a built-from-scratch model doesn’t drift into inaccuracy over time.
• Cost: Building from scratch requires a higher up-front investment in development, plus fees for API calls to an LLM. Customizing prebuilt AI agents from a vendor such as Oracle doesn’t incur any charges beyond the subscriptions for SaaS applications that companies already pay.
• High-quality data: Business data needs to be prepared for AI before agents can use it. That often entails transforming the data into vector embeddings, which mathematically show relations among concepts, helpful for inferring users’ intent when they pose questions. Organizations building agents from scratch also need to watch for “overfitting,” in which an LLM hews too closely to the data it was trained on and can’t generalize to new fields of knowledge.
• Governance and oversight: Businesses may want agents that can document their work and be transparent to line-of-business managers who aren’t IT pros. IT departments also may want to consider that agents don’t have access to sensitive data that isn’t supposed to be made public or disclosed to certain employees.

2. Select an LLM—or get one out of the box. SaaS application vendors that enable their customers to refine agents in a design studio will likely preselect which LLMs their software will interact with, or give admins a limited choice. Organizations building from scratch will need to choose from LLMs from the likes of Anthropic, Cohere, Google, IBM, Meta (developer of the popular Llama models), Microsoft, Mistral, and OpenAI. This approach can give those businesses control over all layers of their agentic software stack, including the underlying model. It also means they’re responsible for maintaining many more software components compared with customizing off-the-shelf agents.

3. Design a workflow and define the tools. Even tailoring prebuilt agents is a job for an applications administrator, not a general business user. Admins can start with predesigned workflow templates—use cases with code behind them in a catalog view—or create new, customized workflows. To define prebuilt agents’ workflows, admins type specific, natural language instructions into fields in an agent design studio or select actions from lists to specify how the agent should interact with users, display data, or schedule appointments. Admins can also choose which tools the agent should use to answer questions, and they can provide sample questions employees might ask.

This process helps define the agent’s role, describing in plain terms how it should carry out a job and what information it will need to access. For example, an agent within an HR application that helps explain health benefits to employees will need access to medical, vision, dental, and other healthcare policy documents, whereas a financial benefits agent may need to tap information on employer-sponsored retirement and stock plans (more on that below).

4. Upload documents for RAG. Now that the agent has its instructions and tools, an admin can use a documents uploader to prepare company documents for retrieval augmented generation (RAG)—an AI technique that supplies an LLM with business documents and data at runtime to augment what the model learned during its training. The administrator provides natural language instructions on how the agent should use the documents. Effective agent builder software abstracts away the vector database that helps deliver highly relevant results at runtime based on what a computer user intends to find.

5. Click to create. Having laid the foundation with instructions, topics, and documents, the admin can create an agent in a design studio simply by naming it and clicking a UI button. Natural language instructions let the workflow (or other agents) understand its capabilities. As they’re running, AI agents are designed to learn how to improve their performance through a mathematical trial, error, and reward process called reinforcement learning.

Companies building from scratch without a design studio may need to add integrations to financial, HR, customer management, and other applications, as well as users’ databases and documents. AI agent frameworks provide an alternative to writing code from scratch by providing software architectures, communication protocols, connectors to cloud and local data sources, and monitoring tools to help businesses build new agents. Popular open source frameworks include LangChain, LlamaIndex, and Microsoft Research’s AutoGen.

Agent studio environments can also include a framework under the hood that admins don’t need to access directly.

6. Set boundaries. Now it’s time to put up guardrails to help ensure that agents retain their accuracy and can identify when to seek approval before carrying out actions. The admin setting up the agent can, for example, add a requirement to get approval from staff before sending an email or updating a record.

Admins can also set conditions under which a question can be answered, or they can add instructions that require the underlying LLM to either pull information from a company IT system or ask the user for clarification, instead of inventing an answer (a drawback of generative AI called hallucinating). For instance, an admin can type: Make sure you have information regarding the number of dependents either by asking the user or querying the system. If you do not know the answer, do not make up a response.

Agents can also be designed to inherit content moderation capabilities from the cloud service on which they’re running.

7. Test, deploy, and monitor. Through a test area in the studio, admins can run through a sample interaction to gauge whether the agent’s responses are helpful and relevant, and check which sources it cites. They can also see how a user interaction would change if the organization altered the agent’s instructions or its underlying LLM. Then an admin can deploy the agent from right in the design studio.

Agents can improve their performance over time by measuring which combinations of RAG data and user prompts yielded the most useful outcomes. Business managers can then rate agents’ performance, incorporating the feedback into future interactions with users.

Use Oracle AI Agent Studio for Fusion Applications to Customize Agents in Oracle Applications

Oracle AI Agent Studio lets IT admins set up AI agents in Oracle Fusion Cloud Applications that are designed to assist users with a variety of tasks, including calling up their paid-time-off balances, pulling up customers’ purchase histories, processing product returns, and analyzing manufacturing equipment photos to estimate the cost of repairs.

Fusion admins start with prebuilt templates, which appear as tiles in their workspace and contain the necessary code to get started. Agent designers then instruct the agent they want to deploy about the scope and limits of its function, and which documents and other data sources it needs to seek out for information. They can also create new agents from scratch. The agents are included in customers’ Fusion subscriptions at no additional cost.

How to Create an AI Agent FAQs

What does an AI agent do?

AI agents are virtual assistants deployed inside of business applications or personal productivity software to help address computer users’ questions or help them complete tasks. Unlike earlier software assistants, which relied on precoded rules and workflows, AI agents are designed to understand natural language prompts and context, while adapting to new situations.

Are AI agents the future?

AI agents could become increasingly more useful as they are deployed across different business applications with less reliance on human intervention, and as they learn from interactions with more business users and consumers over time.