AI agents are being hyped as the future of automation, but if you look at real-world examples, you might wonder: haven’t we been using these for years? Spam filters, recommendation systems, and personal assistants have been around for decades.
The difference isn’t just the technology itself but how these systems are evolving. AI agents today don’t just follow static rules - they reason, adapt, and even collaborate. This article goes beyond the theory to explore how AI agents are actually working in the real world right now.
1. Simple reflex agents
Simple reflex agents take an input and perform an action using a rule/rules. A good example is Microsoft Excel’s “Conditional Formatting” agent: "If a cell’s value is less than 50, highlight it in Red", "If a cell contains the term ‘complain,’ call the email function," etc. These agents work without any reasoning or memory; this simplistic design makes them uncomplicated and fast, finding various applications, like:
Autocorrect
Most of the autocorrect systems work on a simple principle: compare the typed word with the given dictionary. If it’s not a match, underline it in red and/or suggest a replacement.
Classic Eliza chatbot
I remember being fascinated by this chatbot back in the early 2,000s. Having no idea of how quickly these chatbots will advance in a couple of decades, combined with relatively calmer days of computing made it quite an interesting agent. But it was limited as it had neither reasoning nor history capabilities: Even if you ask it the same question repeatedly, it would repeat it the same way every time. For example, this is a snippet of a chat I had with Eliza’s libre version.
Spam filters
Some basic email/sms spam filters take just basic phrases and mark them as spam. “Congratulations! You won a prize,” or “Your credit card is hacked,” etc, can land an email directly into the “Junk” folder.
2. Model-based reflex agents
Some reflex agents use an internal model, too, like the spam filters in modern email software. These models are more flexible than normal reflexive agents and can use history/context too.
Hawk-Eye system
Hawk-Eye is a famous system commonly used in Cricket and some other sports (like Tennis and Volleyball) for ball tracking. It's not only reflexive but also uses an internal model for ball projection.
CheerioCrawler
Any web crawler using some model – no matter how basic – is a good example of a model-based reflex agent. For example, a basic crawler using BeautifulSoup or CheerioCrawler, which uses Cheerio, a fast HTML parser, to extract relevant data.
3. Goal-based agents
Goal-based agents aim to achieve a specific objective, such as driving an autonomous vehicle or winning a computer game. Unlike reflex agents, these goal-based agents use some reasoning to achieve their goals.
Actors
On the Apify platform, you can build AI agents for specific objectives. For example, a backlink-building agent for SEO or a finance monitoring agent. These agents are Actors that can use a combination of web scrapers and some reasoning/processing to fetch and compile data.
Fraud detection systems
Fraud detection systems analyze transaction patterns and can flag or block suspicious activities, like drawing a handsome amount in consecutive transactions over a short time, or from some other geographic location, etc.
Virtual PAs
These virtual personal assistants like Siri, Alexa, Google Assistant are aimed at mimicking a personal assistant and help users with some little tasks like setting reminders, controlling smart devices, helping with emails, etc.
4. Utility-based agents
While goal-based agents are aimed at achieving the objective, utility-based agents consider several factors before undertaking the most optimized action. For example, the Autopilot system always considers factors like arrival time, turbulence avoidance, fuel mileage/efficiency, airspace(s) to avoid, etc. while navigating the plane. Since we are always constrained by some factors in the real world, we can find many examples.
Autonomous vehicles
Autonomous vehicles can vary in complexity, but they account for different factors like fuel mileage, traffic congestion, etc. (similar to Autopilot).
Load-balancers in cloud computing
AWS, Azure, and other cloud computing platforms use agents (known as load balancers) to distribute traffic across servers and avoid congestion as much as possible.
Recommendation systems
All famous social media and streaming platforms use recommendation systems, which are essential goal-based agents (ensuring users don’t go away) but also involve utility considerations, like avoiding content fatigue (by variety) and balancing popular content and personalized discovery.
5. Learning agents
While model-based agents include a reasoning module, it isn’t dynamic (in a learning sense). Learning agents use some learning algorithms (mostly reinforcement learning). For example, we don’t need to go beyond ChatGPT as it continues to learn based on user interactions. Many of the agents we discussed can be learning agents, too, like the Netflix recommender system, which evolves with our use/feedback, some advanced fraud detection systems that continue to evolve based on new types of attacks, and so on.
6. Multi-agent systems
There are instances where we need multiple agents to complete a task. For example, a railway ticketing system may have an agent for the ticket sale and another one checking for the planned strikes. A multi-agent system can have different types of architectures, including centralized and decentralized.
OpenAI Five
This multiagent system made a lot of headlines in 2018 (roughly the same time when GPT-1 was introduced) by beating the top players in the famous Dota 2 game. In Five, different agents collaborate to compete vs. human players in real time.
NASA’s Mars rovers
NASA’s Perseverance Rover is another classical example of a multi-agent system in which a rover and satellites work together to search for signs of ancient microbial life.
7. Hierarchical agents
Hierarchical agents are systems structured in multiple layers, which work in a top-down approach. For example, in an industrial automation system for car manufacturing, higher layers will manage the overall production, while lower layers will control sections or individual machines (assembly line, paint, specific quality checks, etc). Hierarchical agents are usually complex, and while they are commonly found in industries, we can find some other examples too. Hierarchical NPC behavior is common in the games: higher layers would determine mission objectives, while lower layers would choose the navigation paths to achieve the mission. Some advanced LLM agents, like AutoGPT, are also hierarchical agents.
Conclusion
With an increased interest in AI agents, it helps to know the different agent types and real-world examples. Agents are no longer just theoretical concepts - they're actively shaping industries and everyday experiences. From goal-based assistants to learning-driven systems, these agents are powering everything from self-driving cars to advanced web automation.
More about AI agents
- How to build an AI agent - A complete step-by-step guide to creating, publishing, and monetizing AI agents on the Apify platform.
- AI agent workflow - building an agent to query Apify datasets - Learn how to extract insights from datasets using simple natural language queries without deep SQL knowledge or external data exports.
- 11 AI agent use cases (on Apify) - 10 practical applications for AI agents, plus one meta-use case that hints at the future of agentic systems.
- What are AI agents? - The Apify platform is turning the potential of AI agents into practical solutions.
- AI agent architecture in 1,000 words - A comprehensive overview of AI agents' core components and architectural types.
- LLM agents: all you need to know in 2025 - LLM agents are changing how we approach AI by enabling interaction with external sources and reasoning through complex tasks.
- 7 types of AI agents you should know about - What defines an AI agent? We go through the agent spectrum, from simple reflexive systems to adaptive multi-agent networks.
