What you need to know about developing AI agents

Imagine developing an API for the generative AI era. This API would support natural language input, leverage intelligence from large language models, automate decisions by integrating with SaaS and enterprise systems, and enable business process orchestration by connecting to other genAI-enabled APIs.

That’s one way to understand what AI agents are and how they work. They integrate the reasoning of RAG-enabled LLMs with access to contextual business knowledge and the ability to act semi-independently. Whereas an LLM takes natural language input and answers questions, an AI agent is more like a business partner that actually gets work done. The more sophisticated autonomous agents are designed to work largely free of human interactions. Being autonomous, they incur more risks in interpreting inputs and taking actions.

AI agents are transforming the future of work by providing developers, business users, and others with a role-based partner that proactively automates tasks and acts as a knowledgeable collaborator. AI agents are generating significant business interest for their range of use cases and applicability beyond just intelligent information processing with genAI. Many AI agent use cases are not moonshots but solid proposals likely to get funded.

“AI agents have evolved from rule-based systems to intelligent, genAI-driven tools capable of natural language interaction,” says Ram Palaniappan, CTO of TEKsystems. They can be deployed in diverse use cases, such as procurement to evaluate suppliers and place orders, customer support to perform complex troubleshooting, and healthcare to provide diagnosis summaries. Organizations with process-driven, human-intensive operations benefit from scalability, 24/7 availability, and streamlined workflows.”

AI agents versus APIs

Let’s start by understanding how AI agents differ from APIs and other web services.

AI agents accept natural language and other non-technical inputs, including emerging AI agents that will accept voice, images, and video. Contrast this to APIs that only accept system-oriented inputs like JSON and XML.

Instead of coded business rules, AI agents connect to RAG, language models, and other genAI models for relevant knowledge.

AI agents can reason the next courses of action based on their assigned role and the defined guardrails of what they are not supposed to do. Because APIs are rules-driven, they require far more programming time and effort to simulate different roles, decisions, and actions.

Both APIs and AI agents can perform transactions and communicate with other services. AI agents can also act on their decisions and send natural language prompts to other AI agents.

AI agents provide natural language outputs to human collaborators. Non-technical people can validate an AI agent’s level of understanding and its actions because the decision-making steps and connection to APIs and other AI agents are summarized in natural language.

“Developers building AI agents for use cases like customer service should use natural language to encode business logic instead of code,” says Deon Nicholas, co-founder of Forethought. “They should equip agents with the ability to communicate with APIs in the same way humans communicate with websites. This will unlock truly agentic AI, which can take action and resolve issues, delivering a true value add.”

What kind of AI agent do you need?

There are several types of AI agents, classified by how they make decisions and perform actions. Model-based agents replace rules with AI models and supporting data, while goal and utility-based agents compare different scenarios before selecting a course of action. The more sophisticated AI learning agents use feedback loops to improve results, while hierarchical agents work in a group to deconstruct complex tasks.

Can you wrap an API with a natural language interface and call it an AI agent? The answer is yes; these are simple reflex agents that leverage rules to connect natural language input to action.

What are the prerequisites to developing AI agents?

When developing AI agents, you should be aware of prerequisites involving platforms, data, integration, security, and compliance.

“The success of AI agents requires a foundational platform to handle data integration, effective process automation, and unstructured data management,” says Rich Waldron, co-founder and CEO of Tray.ai. “AI agents can be architected to align with strict data policies and security protocols, which makes them effective for IT teams to drive productivity gains while ensuring compliance.”

Mike Connell, COO of Enthought, says you need a high volume of clean and (for some applications) labeled data that accurately represents the problem domain to train and validate models. Connell says, “A robust data pipeline is essential for preprocessing, transforming, and ensuring the availability of real-time data streams to refine the model and keep it calibrated to a changing world. Additionally, you should consider the need for domain-specific ontologies or embeddings to enhance the agent’s contextual understanding and decision-making capabilities.”

Regarding security and compliance, Joseph Regensburger, VP of research at Immuta, says AI agents have identities, so access to complex AI chains and knowledge graphs requires controls as if they were human. Regensburger recommends, “Capture the frequent changes in regulations and business agreements in an access control solution and enforce them on all potential human and machine actors.” Keeping up with changing business rules is essential to ensure AI agents are not developed based on outdated usage agreements.

Technologies and platforms for developing AI agents

Enterprise platforms such as Appian, Atlassian, Cisco Webex, Cloudera, Pega, Salesforce, SAP, ServiceNow, and Workday have announced AI agent capabilities embedded in their workflows and user experiences. For example, the Workday recruiter agent helps HR recruiters find and hire talent, while Atlassian’s AI-powered virtual service agent helps automate tier-1 support issues.

Some platforms also have capabilities for subject matter experts and non-technical business users to develop their own AI agents. Salesforce Agent Builder allows non-technical users to create customized AI agents. Users describe the agent’s role and select topics representing the work to be done, and the AI maps these to activities that can be performed on the platform. Other platforms with AI agent-building capabilities include Cisco Webex AI Agent Studio, ServiceNow Agentic AI, and Tray.ai Merlin Agent Builder.

One option for AI agent development comes directly as a service from platform vendors that use your data to enable agent analysis, then provide the APIs to perform transactions. A second option is from low-code or no-code, automation, and data fabric platforms that can offer general-purpose tools for agent development.

“A mix of low-code and pro-code tools will be used to build agents, but low-code will dominate since business analysts will be empowered to build their own solutions,” says David Brooks, SVP of Evangelism at Copado. “This will benefit the business through rapid iteration of agents that address critical business needs. Pro coders will use AI agents to build services and integrations that provide agency.”

A third option comes from developing agents natively with code, AI agent builders, or LLM application frameworks.

“You can either build AI agents natively—such as with Python or C++—or use a framework like AutoGen, LangGraph, or CrewAI, but these may not scale well or have sufficient guardrails,” says Abhi Maheshwari, CEO of Aisera.” You also need modern data infrastructure, such as a data lake or lake house. Data must also be relevant for the domain and “seamlessly integrated using techniques like fine-tuning of LLMs or RAG,” says Maheshwari.

Organizations looking to be early adopters in developing AI agents will likely need to review their data management platforms, development tools, and smarter devops processes to enable developing and deploying agents at scale.

“To accelerate agent development, companies will need a robust set of tools that allow them to design, customize, deploy, and monitor agents at scale,” says Maryam Ashoori, director of product management of watsonx.ai, IBM. “This includes models optimized for function calling, middleware to orchestrate agents and connect them with broader enterprise toolsets, optimized runtime, technical guardrails, and governance capabilities to ensure they operate as intended. It will also require tooling that caters to a wide set of users and skillsets, from pro-code tools for developers to low-code and no-code tools for business users to embed them in daily workflows.”

Testing AI agents

Testing LLMs and validating accuracy requires human testers, automation, and synthetic data for basic accuracy testing, while more sophisticated techniques leverage secondary AI models and use generative adversarial networks (GANs) to test at scale.

Rahul Pradhan, VP of product and strategy, AI, data, and analytics at Couchbase, says, “Testing for accuracy via sophisticated observability tools, feedback loops, and fallback mechanisms will help organizations establish trust in AI agents, marking a leap toward leveraging agents that can perform tasks with autonomy.”

Mike Finley, CTO and co-founder of AnswerRocket, says AI agents can be tested for accuracy in two stages:

Require AI Agents to provide documented proof points where any facts used or quoted include their sources and any decisions made include documented logical steps describing their inputs.

AI verifiers are supervisor agents whose job is to watch the work of other AI agents and review accuracy while looking for subtle cues like a shift in tone.

AI agents will redefine the workforce

LLMs and RAGs received significant hype on generative AI that’s given way to the potential of how AI agents can impact productivity across a wide range of business workflows. As more platforms make agents available and development platforms scalable, new human and AI responsibilities will likely emerge.

“Agentic AI will reshape the workplace and create new roles, such as ‘Agent Managers’ to oversee specialized agents, strategically guide these systems, and ensure alignment with business roles, similar to supervisors managing teams today,” says Artem Kroupenev, VP of strategy at Augury. “As multi-agent systems grow, HR-like departments may emerge to manage a hybrid workforce of human and AI agents, focusing on training, coordination, and performance metrics. This hybrid approach could blend human intuition with machine efficiency for better productivity.”

The key to growth may not be in how easy it is to develop AI agents, but in whether and how organizations will trust them and whether employees will embrace their capabilities.