What are AI Agents?

Introduction

In a relatable scene from the television show Malcolm in the Middle, Hal attempts to replace a lightbulb, entering a rabbit hole of dependencies to achieve his original goal. This is the nature of human action: for every task there are many considerations and related tasks. As Hal's character shows, modern life is complicated, but Artificial Intelligence (AI) agents can be our assistants to help with everything from home improvement to travel booking and more.

AI agents manage life's complexities especially well, moving beyond the limitations of chatbots to take actions on a user's behalf. Agents manage inputs, reason about the data, and use a variety of tools to achieve a goal. AI agents are not new; assistants like Siri and Alexa and self-driving cars – a form of AI agent – have been around for years. What makes AI agents most relevant today is that LLMs are unlocking more capabilities for non-developers to both interact with and create AI agents.

What is an AI Agent?

An AI agent is an autonomous, non-deterministic program that can gather input, reason about collected data, and act on the knowledge it has in order to achieve a goal. Unlike AI chatbots, AI agents can take actions, like placing an order or booking a reservation.

Characteristics of AI Agents include:

• Autonomous: operate independently without human guidance.
• Action-oriented: independently execute different tools.
• Interface agnostic: input and output beyond chat (voice or text) based systems.
• Non-deterministic: dynamic actors that use reasoning and can learn from their inputs and adapt to new information.

AI agents are ideal assistants for handling complex problems. This means that agents must be well tested, provide transparency, and work within defined boundaries.

How AI Agents Work

AI agents start with some form of user defined goal. This goal could come from human interaction, like a customer calling into a support phone number for help returning their order.

While a human will set the goal and can be involved to approve certain decisions, AI agents act independently to decide which data to parse, how to interpret that data, and which actions to take in order to achieve the goal. Agents generally use an LLM, or Large Language Model, under the hood to perform these functions.

Typical steps of an AI agent include:

1. Define a goal with user input
2. Retrieve information and collect knowledge about the task
3. Reason about and parse collected information to create a plan
4. Take action via tools, for example search, use calculators, make database queries, perform translation, or make an API call
5. Use memory about the person to personalize the interaction. This could include previous chat logs, purchase history, or other account information.

These steps are repeated until the goal is achieved. AI agents are generally considered autonomous, but human-in-the-loop approvals are common to validate the progress of an agent. Designing agents to provide transparency and allow for human intervention and feedback is critical for building trust in the systems.

Where AI Agents Shine

AI agents are a natural fit for problems with open-ended inputs like customer support: while the majority of your support calls are likely about a small number of things, the way a customer could ask for those things is infinite. LLMs are especially good at determining intent without needing to write complicated, ballooning if-else statements. This allows an AI agent to replace parts of your IVR to handle more mundane, repetitive tasks like checking an order status.

Agents can also provide increased efficiency and availability, providing 24/7 support for a subset of problems. Agents can also personalize interactions, using previous interactions and customer profiles to provide context and memory to conversations. For example, an AI agent could know to add delivery instructions not to ring your doorbell to avoid upsetting your dog.

Agents can also learn and adapt from dynamic sources, so you don't have to reprogram your IVR options with every new policy. The best part is that you can design AI agents to hand off to a human for sensitive or more complicated tasks.

Challenges and Considerations for AI Agents

Like any software system, robust and trusted AI agents start with good design. Anthropic's core principles for building AI agents include simplicity, transparency, documentation, and testing. Because AI agents provide non-deterministic outcomes, solid testing and the option for human intervention are important when building robust and trustworthy agents. Use human-in-the-loop interactions to periodically make sure you're on the right track.

While AI agents may mitigate some of the social engineering threats of traditional call centers, data privacy and security are still important considerations. Build secure systems to prevent threats like prompt injection, information disclosure, and more. Ensure that users are authenticated before an agent can take action and that tools like APIs are properly secured.

How Twilio can help you build AI Agents

AI agents move beyond the chatbot interface into the systems we use every day. That includes interfacing with other communication channels like Voice and Email or as integrated software in your existing applications.

• AI Assistants is an opinionated platform to build and host conversational AI Assistants for your customer-facing use cases. Iterate quickly with an easy to use interface for connecting knowledge sources, tools, and customer data.
• ConversationRelay enhances voice interactions by integrating real-time speech recognition and synthesis using an LLM of your choosing.
• Twilio Segment provides customer memory for building personalized interactions.
• Seamless hand off via our omnichannel communications platform using built-in tools like Twilio Studio.
• Twilio Functions is a serverless environment that allows you to quickly create & deploy production-grade functions an AI Agent can call via a tool.

AI agents are incredibly promising tools and we're excited to help you build this next iteration of customer engagement. Twilio is optimistic about the future of AI and understands that the best customer experience still provides the human touch when necessary – and we're excited to connect you to your customers at every step of the journey.