Case study
AI Meets Rugby: Framesports automates match analysis on Ori Serverless Kubernetes
Learn how Ori Serverless Kubernetes is helping Framesports analyze Rugby matches with AI.
Agentic AI is the next frontier in AI adoption. Discover more about AI agents in this blog post: what are they, types of agents, benefits, AI agents vs LLMs, and how agents can transform business workflows.
As the artificial intelligence (AI) landscape evolves rapidly, driven by large language models (LLMs) and generative AI, businesses are eager to embed AI deeper into their workflows. According to a Redpoint Global survey, 73% of consumers see potential for AI to impact customer experience (CX), especially in digital settings. Nearly half (48%) of all consumers say they would interact with AI more frequently if it would enhance their experience. This is where Agentic AI shines, making generative AI more actionable, contextual, and autonomous. Read on to know about AI agents, their impact, architecture, and more.
What are AI agents?
AI agents are autonomous systems that interact with users and its environment to make decisions and complete tasks. These agents use LLMs or other AI/ML models, and contextual understanding to perform a wide range of functions, from customer service, data analysis, autonomous driving, smart home management, and much more. Many AI agents can learn from experience and adapt to new situations, making them valuable in several industries for improving efficiency, productivity, and user experience.
Agentic AI supercharges your business workflows
Agentic AI has the potential to transform businesses across industries in several ways:
- Automation of Complex Tasks: AI agents can handle sophisticated tasks requiring nuanced decision-making, strategic planning, and flexibility. Industries such as finance, healthcare, and transportation are seeing significant improvements through AI-driven analysis, diagnostics, and autonomous operations.
- Amplified Efficiency: By automating both routine and complex tasks, agentic AI allows human professionals to focus on creative, strategic, and higher-value activities, leading to increased productivity and efficiency across sectors.
-
Better User Experience: In customer service and user interaction scenarios, agentic AI can provide personalized, context-aware responses, leading to higher user satisfaction and engagement.
-
Innovative Applications: From smart homes and creative services to advanced robotics and autonomous vehicles, AI agents can be employed in a wide range of use cases and in many cases driving innovative new applications.
How are AI agents different from LLMs
While AI agents and LLMs share some similarities, they differ in several important ways:
|
AI Agents
|
LLMs
|
|
|
Autonomy
|
Act autonomously, taking actions and making decisions on their own.
|
Passively generate responses based on input text.
|
|
Context
|
Use short-term and long-term memory to maintain relevant context, generating accurate responses and actions.
|
Generate text based on patterns from large datasets but don’t truly understand context.
|
|
Iteration
|
Reflect on their work and improve iteratively, similar to processes in business settings.
For example, creating a marketing blog post needs research, an outline, a first draft, multiple iterations based on further research, reviews from colleagues and so on. Agents can be modeled after this behavior to make sure their output is more precise and meets the standards for business processes.
|
LLMs rely on pre-existing knowledge and natural language processing, however they’re unable to evaluate their output and improve on the go.
Commonly used with zero-shot prompts, LLMs generate their output token by token and don’t perform an iterative process of output generation and decision-making .
|
|
Task execution
|
Integrate various tools and modules to perform a wide range of tasks, from scheduling appointments to smart home management and customer support.
|
Excel at generating human-like text and answering questions but are limited in executing tasks.
|
|
Real Time Interaction
|
Access the Internet, databases, and APIs for up-to-date information, adjusting responses in real time.
|
Use training data that might not be up to date despite advances in retrieval-augmented generation (RAG) enhancing output quality.
|
|
Mix-and-match
|
Create best-of-breed workflows with multiple LLMs or other types of AI models at their core, offering scalability across applications and industries.
|
Although Mixture of Experts (MoE) LLMs are driving a more modular LLM architecture, they’re still limited in adapting to diverse workflows requiring varied knowledge sources and tools.
|
Types of AI agents
.png?width=1920&height=960&name=Type%20of%20AI%20Agents%20(2).png)
Simple Reflex Agents: They have no memory of past actions or events. Simple reflex agents respond to their perception of their current environment and make decisions based on predefined rules or learned patterns, akin to a thermostat turning the HVAC system on based on temperature settings.
Model-based Reflex Agents: These agents are more sophisticated than simple reflex agents and create a representative model of their world to make decisions instead of relying just on rules. They use this model to make decisions by considering the history of interactions and predicting the outcomes of their actions.
Goal-based Agents: are driven by specific goals and they choose a line of action based on whether it can help achieve the goal. These agents sometimes consider a sequence of actions which lead to the goal, like a chess player planning moves. A physical robot that plans its moves by sequencing multiple steps is an example of a goal based agent.
Utility-based Agents: Similar to goal-based agents, these agents strive to achieve goals, but they aim to maximize a utility function while achieving that goal . An investment advisor agent designed to maximize return on investment (ROI) is an example of a utility based agent.
Learning Agents: These agents improve their performance over time by learning from their experiences. They can adapt to changes in the environment and refine their decision-making processes. For example, a digital ad management agent can optimize its ad placement based on continuous audience feedback.
Multi-agent Systems: mimic the collaborative workflows found in modern workplaces, where multiple reviewers provide feedback or share responsibilities to complete larger projects. These systems thrive on iterative feedback, with agents critiquing each other to refine decisions and actions.
Imagine a team of specialized agents, each an expert in a specific area, working together seamlessly. For instance, a multi-agent design system could feature agents powered by different language models, each excelling in areas like coding, math, or other specialized knowledge. This collaborative approach makes problem-solving more efficient and precise.
Hierarchical Agents: are a type of multi-agent system where decisions are made in multiple levels. Each level handles a different part of problem solving, with higher agents guiding and controlling agents which are lower in the hierarchy. Hierarchical agents are useful in fail safe environments such as manufacturing and regulated industries where oversight of workflows is crucial to prevent undesired outcomes.
Designing AI agents
Based on the type of agent, the AI agentic workflow consists of several modules, the most important of them being the core agent, memory, planning and tools. Other agents are an integral part of the multi-agent architecture whereas the environment can often be an important component of the learning system.
Here’s a short overview of each of these modules:
.png?width=1920&height=960&name=The%20AI%20Native%20Cloud%20(2).png)
Agent Core acts as the brain of an agent, handling key decisions and actions. Usually built around a powerful AI/ML model such as an LLM, it leverages model knowledge to guide how the agent thinks and behaves. The Agent Core knows the agent's goals, provides a "user manual" for its tools, explains how to use planning modules, and adds relevant memory from past conversations based on user queries. It can also include a persona to shape the agent's behavior and response style.
Memory modules are vital for AI agents, acting like a diary of the agent's actions and user interactions. These memories are retrieved based on a mix of relevance, importance, recency, and other factors, ensuring the agent can provide accurate and contextually appropriate responses. There are two types of memory modules for AI agents:
- Short-term memory: Tracks the agent's thought process for answering a single question. This memory could be implemented using a queue/FIFO data structure
- Long-term memory: Keeps a history of interactions over weeks or months. Local storage or a scalable vector database can be used to implement text embeddings or vectors to be used as long-term memory.
Tools: are well-defined workflows that allow agents to perform specific tasks efficiently, much like specialized third-party APIs. For example, an agent might use a RAG pipeline to generate context-aware answers, a code interpreter for complex programming tasks, an API to search the internet for information, or services for tasks such as weather updates or instant messaging or calendar events.
Planning module: Tackling complex problems, like analyzing financial reports to answer layered business questions, requires a nuanced approach. An AI agent handles this by decomposing questions, along with reflection or critique. For instance, answering "What were the main takeaways from an earnings call” involves breaking it down into smaller questions, such as financial results, business environment and growth strategy. Each sub-question can then be further divided, with a specialized AI agent guiding the entire process for clarity and precision.
Actions: involve executing tasks based on cognitive decisions and can be physical actions in real-time such as actuators in case of robots or virtual actions such as calling APIs or updating content, and more.
Users: can be human collaborators or other virtual AI agents who provide inputs ranging from prompts to objectives, deadlines, contextual knowledge and even feedback.
Environment: The agent core can also calibrate its response continuously based on external inputs such as text, camera, audio, and sensors which makes it particularly useful and more precise in physical environments.
Build the future of AI on Ori
Agentic AI is a significant advancement in artificial intelligence by offering autonomy, scalability, and context-awareness. AI agents and assistants are already helping businesses harness the potential of AI. For example, Klarna’s AI-powered customer assistant has reduced issue resolution time from 11 minutes to less than 2 minutes, and driven a 25% reduction in repeat queries.
Ori’s AI Native cloud is purpose-built for AI/ML workloads such as training models, inference-based AI solutions, fine tuning models and much more. Backed by top-notch GPUs, performant storage and AI-ready networking Ori enables AI-focused startups and enterprises to:
- Deploy AI-Optimized GPU instances for training, finetuning and inference workloads.
- Reduce your GPU costs by up to 75% when compared to hyperscale cloud providers.
- Scale effortlessly from on-demand to custom private clouds in a variety of configurations ranging from bare metal and virtual machines to Serverless Kubernetes.
