What Are AI Agents? (Benjamin Labaschin)
- Martin Swartz
- 1 day ago
- 13 min read
INTRODUCTION
AI agents have quietly become one of the most consequential developments in computing history. Benjamin Labaschin places them alongside the invention of the computer, the personal computer, the internet, the smartphone, machine learning, and cloud computing as milestones that thrust society forward. If you are reading this, he writes, you are living during another one of those milestones.
But what exactly are AI agents? The answer is more practical than mystical. AI agents are tools designed to let users interact with large language models (LLMs) to achieve a more productive or creative workflow as seamlessly as possible. Before AI agents existed, anyone wanting to use AI had to build their own statistical language models: a time-consuming, technical, and expensive endeavor. Now, users simply log in to an interface and conduct business ranging from asking questions of their documents to getting help with their homework.
This Book Essential is for anyone who has heard the term "AI agent" and wanted a clear, grounded explanation of what these tools are, how they work, when to use them, and what risks they carry. Labaschin wrote this report for O'Reilly Media to give business people, developers, students, and curious readers a starting point. He uses a fictional company called Very Fake Company, Inc. to walk readers through three categories of AI agents: document question-and-answer agents, chatbot agents, and code assistant agents.
The book also looks forward. Labaschin identifies augmented search, product integration, infrastructure costs, and security legislation as the frontiers that will shape the next phase of AI agent development. Whether you are building agents, buying them, or simply trying to understand the conversation, this Book Essential distills the core ideas into an actionable format.
U365'S VALUE PROPOSITION
WHO THIS IS FOR
Business professionals who want to understand what AI agents are before investing time or budget in them, and need a clear non-technical explanation.
Software developers and engineers who are evaluating AI agent tools like GitHub Copilot, Chainlit, or ChatGPT for their daily workflow.
Students and self-learners who are encountering AI tools for the first time and want a structured introduction to the landscape.
Managers and decision-makers who need to understand the risks, costs, and infrastructure implications of deploying AI agents at scale.
Anyone curious about AI who has heard terms like LLM, agent, and prompt but wants a grounded explanation without hype.
KEY TENSIONS
Simplicity versus power: AI agents abstract away the complexity of LLMs, making powerful models accessible through simple text interfaces. But that same simplicity can mask the technical limitations and failure modes of the underlying models, leading users to trust output they should verify.
Open source versus proprietary: The book highlights both proprietary agents like ChatGPT and open source alternatives like Meta's LLaMa 2 paired with HuggingFace. Open source gives control and customization, while proprietary tools offer polish and ease of use at a subscription cost.
Cost versus capability: Running AI agents can be computationally expensive. Labaschin notes that maintaining the LLM engines that power agents requires significant infrastructure, though the landscape is shifting rapidly with local models and streaming approaches reducing costs.
Adoption gap: Only a quarter of companies with more than 250 employees have invested in AI of any kind, compared to 7.7 percent of firms with fewer than 10 employees. This imbalance threatens to widen the gap between large "superstar" firms and smaller companies.
Utility versus risk: The same properties that make AI agents useful also make them dangerous. Users have fed confidential company and personal information to agents, compromising intellectual property. LLMs trained on materials without permission raise legal and ethical questions that legislators are still grappling with.
Human oversight versus automation: Code assistant agents can generate wrong or misguided output that causes problems to snowball. Labaschin stresses that agents are tools, and users must know when and how to wield them, verifying output with human oversight.
WHY IT MATTERS NOW
AI agents went from research curiosity to mainstream tool in under a year. ChatGPT alone brought AI agents to millions of users. Companies like Alphabet, Microsoft, Meta, Amazon, and Apple have all felt pressure to integrate AI agents into their workflows, and upstart companies are forming overnight to fill gaps the technology creates.
The pace of change makes it difficult to keep up. New agent types, new use cases, and new infrastructure approaches emerge almost daily. A grounded understanding of what AI agents are and how they relate to LLMs gives you a stable mental model you can apply to any new tool that appears.
The risks are also urgent. Security, intellectual property, and ethical concerns are not future problems: they are present problems. Legislators in many countries are already considering laws to limit or prevent the use of AI agents. Understanding the landscape now prepares you to navigate whatever regulatory and technical changes come next.
OVERVIEW
"What Are AI Agents?" is a concise O'Reilly report that explains what AI agents are, how they relate to LLMs, when to use them, and what tools are available today. Labaschin writes in a deliberately accessible style, using a fictional company called Very Fake Company, Inc. to ground each concept in a concrete scenario. The result is a short book that reads more like a guided tour than a technical manual.
The core analogy Labaschin uses throughout is the car and engine metaphor. AI agents are cars: user-friendly interfaces that let you drive. LLMs are engines: the powerful models that actually do the work. Just as you do not need to understand how an engine works to drive a car, you do not need to understand how an LLM works to use an AI agent. But just as some engines suit specific use cases better than others, some LLMs are built for specific purposes like code assistance or document analysis.
The book covers three categories of AI agents in depth: document Q&A agents, chatbot agents, and code assistant agents. For each, Labaschin provides a scenario at Very Fake Company, Inc., then maps the scenario to real-world tools: Chainlit with PandasAI for document Q&A, HuggingFace Chat with LLaMa 2 for chatbots, and OpenAI ChatGPT for code assistance. The final section looks forward at augmented search, product integration, infrastructure, and security legislation.
KEY IDEAS
AI AGENTS FUNDAMENTALS
AI agents as UI wrappers: Labaschin defines AI agents as user-friendly frontends that make interacting with LLMs easier. They focus and limit how users interact with the model, abstracting away technical complexity. ChatGPT, for example, abstracts away the massive complexity of GPT-3.5 Turbo or GPT-4 and lets users interact via simple text.
The car and engine analogy: AI agents are cars, LLMs are their engines. Cars cannot move without engines, and some engines are designed for specific use cases like ATV engines versus school bus engines. Similarly, AI agents cannot function without LLMs, and some agents work better with particular LLMs suited to their purpose. Users do not need to understand how the engine works to drive the car.
LLMs as the brains: LLMs are the "AI" in AI agents. They are trained to extrapolate and interpret natural language text in the context you provide. Some are better at math, others at jokes or Spanish, and some are general purpose. The push to create AI agents came from the recognition that for LLMs to become ubiquitous and beneficial, user-friendly interfaces had to be developed.
Document Q&A agents: These tools solve the problem of information glut. They ingest documents, store them in memory, and consolidate and regurgitate relevant information through a natural language interface. Users simply ask questions and receive answers without searching through documents manually. Labaschin demonstrates this with an onboarding scenario at Very Fake Company, Inc.
Chatbot agents: More dynamic than document Q&A agents, chatbot agents are fed prompts ahead of time before users interact with them. Prompts are instructions the agent follows at all times. Chatbot agents can connect to third-party systems like databases and the internet, communicate in multiple languages, cope with difficult users, and operate around the clock.
Code assistant agents: These agents are fueled by models designed specifically to help users write code. Popular examples include GitHub Copilot, Amazon CodeWhisperer, and Hugging Face StarCoder. They edit error-ridden code, autocomplete simple functions, and design templates for difficult coding problems, reducing costly context switching.
The importance of explicit prompting: Labaschin shows that the most useful responses from chatbot agents come when users are explicit with their words. In the HuggingFace Chat demonstration, the agent is told exactly the role it should play and asked to explain why LLMs are powerful. Being explicit with instructions lets the agent structure its response into more useful output.
Human oversight is non-negotiable: Code assistant agents can generate wrong or misguided output that causes problems to snowball. Labaschin stresses that users must verify agent output with human oversight and ensure agents are tools that work for you, not the other way around. Understanding how to use these tools is the best way to achieve this.
Real-world tools mapped to agent categories: The book grounds each agent type in real software. Document Q&A uses Chainlit and PandasAI. Chatbot agents use HuggingFace Chat with LLaMa 2 and Streamlit. Code assistance uses OpenAI ChatGPT with GPT-4 or GPT-3.5. Each tool is demonstrated with the same book dataset about computer science titles.
Augmented search as a frontier: AI agents can consolidate and reproduce relevant information, making them natural augmentations or replacements for search engines. Microsoft released Bing AI and Google released Bard. But the pay-per-click ad model that funds search engines may not survive when users can get summarized content without clicking ads.
Infrastructure and LLMOps: Running AI agents is computationally expensive due to the LLM engines. But the landscape is shifting: LLMs like LLaMa 2 can now run locally, and streaming approaches are reducing costs. Labaschin identifies LLMOps as an emerging specialization where experts are likely to be well employed and remunerated.
Security, legislation, and ethics: Users have fed agents confidential information compromising intellectual property. LLMs trained on materials without permission raise IP questions. The default tone and voice of agents carry ethical implications. Legislators in many countries are already considering laws to limit or prevent AI agent use.
SUMMARY
Chapter 1: What Are AI Agents?
The opening chapter defines AI agents as tools designed to let users interact with LLMs for more productive or creative workflows. Before AI agents, users had to build their own statistical language models, which was time-consuming, technical, and expensive. Now users simply log in to an interface and start working. Labaschin introduces the car and engine analogy: AI agents are cars, LLMs are their engines. You do not need to understand how an engine works to drive a car, and you do not need to understand how an LLM works to use an AI agent. The chapter explains that some LLMs are better at specific tasks like math, jokes, or languages, while others are general purpose. The push to build AI agents came from the recognition that LLMs needed user-friendly interfaces to become ubiquitous and beneficial.
Chapter 2: When to Use AI Agents
This chapter introduces three categories of AI agents through scenarios at Very Fake Company, Inc. Document Q&A agents solve the problem of information glut: a new employee with a stack of onboarding documents can upload them and ask questions instead of reading everything. Chatbot agents are more dynamic: they are fed prompts ahead of time that define their behavior, and they can connect to databases and the internet. The chapter demonstrates a chatbot called BERTA that verifies employees and activates software access. Code assistant agents help users write code more productively: the chapter shows a Python-to-Java code translation using a code completion tool embedded in an editor. Each scenario is concrete and uses the same fictional company, making the distinctions between agent types clear.
Chapter 3: AI Agent Tools and How to Use Them
Labaschin maps the fictional scenarios to real-world tools. For document Q&A, he demonstrates Chainlit paired with PandasAI, showing how users can upload a CSV of computer science books and query it for data science titles filtered by publisher. For chatbot agents, he uses HuggingFace Chat with Meta's LLaMa 2, showing the assistant and user personas and emphasizing that explicit instructions produce better responses. For code assistance, he uses OpenAI ChatGPT with GPT-4, showing how the agent can take a data file and generate a Tufte-styled bar chart of pages per year. Each demonstration uses the same book dataset, making the comparisons between agent types direct and practical.
Chapter 4: Moving Forward with AI Agents
The final chapter looks at four frontiers. Augmented search: AI agents can summarize content without ad clicks, challenging the pay-per-click model that funds search engines. Product integration and isolation: only a quarter of large companies have invested in AI, creating a gap between "superstar" firms and smaller companies. Stanford professor Erik Brynjolfsson warns we are "flying blind into what has been called the fourth industrial revolution." Infrastructure: running agents is expensive, but local LLMs and streaming approaches are reducing costs, and LLMOps is emerging as a well-paid specialization. Security, legislation, and ethics: confidential information leaks, IP violations from training data, gender bias in agent voice, and pending legislation all demand attention. Labaschin concludes that the question is not whether AI agents will be used but how they will be used.
IN PRACTICE
1. Start with a clear use case: Before adopting an AI agent, identify the specific problem you want to solve. Document Q&A for information overload, chatbots for customer-facing interactions, code assistants for programming productivity. Do not adopt an agent because it is trendy; adopt it because it fits a real need.
Action: List three tasks in your current workflow that involve searching documents, answering repetitive questions, or writing boilerplate code. Pick one as your first agent pilot.
2. Choose the right LLM engine for the job: Just as ATV engines and school bus engines serve different purposes, different LLMs suit different agents. Code assistance works best with models trained on code. General chatbot tasks can use general-purpose models. Match the engine to the use case.
Action: If your task involves code, evaluate GitHub Copilot or a code-trained model. If your task is conversational, evaluate ChatGPT or Claude. Write down which model fits your pilot task and why.
3. Write explicit prompts: The most useful responses come when you are explicit with your instructions. Tell the agent its role, its constraints, and the format you want. Vague prompts produce vague results.
Action: Draft a system prompt for your pilot agent that defines its role, what it should do, what it should not do, and how it should format responses. Test it and refine based on the output.
4. Verify every output with human oversight: Code assistant agents can generate wrong or misguided output that causes problems to snowball. Treat agent output as a draft, not a final answer. Review, test, and validate before acting on it.
Action: For your first five agent interactions, document what the agent produced and what you checked or corrected. This builds a verification habit.
5. Protect confidential information: Users have fed agents confidential company and personal information that compromised intellectual property. Assume anything you type into an agent could be stored or exposed. Do not paste proprietary code, customer data, or internal documents into public agents.
Action: Review your organization's data classification policy. Identify what information is safe to share with public AI agents and what requires a private or on-premises solution.
6. Watch the infrastructure cost: Running AI agents can be computationally expensive. Track what you spend on subscriptions and API calls. As local LLMs and streaming approaches mature, re-evaluate whether you can reduce costs by switching infrastructure.
Action: For one month, track all AI agent costs: subscriptions, API calls, compute time. Compare this to the time saved to calculate whether the investment is paying off.
7. Monitor the regulatory landscape: Legislators in many countries are considering laws to limit or prevent AI agent use. If your organization depends on AI agents, stay informed about pending legislation that could affect your operations.
Action: Set up a monthly check on AI regulation news for your jurisdiction. Note any proposed legislation that could affect how your organization uses AI agents.
QUOTES
"AI agents are tools designed to allow users to interact with LLMs to achieve a more productive or creative workflow as seamlessly as possible."
"Before AI agents, users would be forced to build their own statistical language models: a time-consuming, technical, and expensive endeavor!"
"At a more granular level, you might think of AI agents as UI wrappers around the models that power them."
"Like cars, AI agents cannot function without their underlying LLM engines. And like cars, some AI agents often work better with particular engines that suit their purposes better."
"Users of AI agents do not need to understand how the car's engine works to drive it. Drivers simply need to know how to use the car itself."
"Document Q&A agents are AI-enabled tools designed for the problem of information glut."
"Chatbot agents are far more dynamic in their capabilities. Whereas the focal point of document Q&A agents is question and answering from source material that you provide them, chatbot agents are fed prompts ahead of time."
"Code assistant agents reduce the need for such costly context switching, allowing for large improvements in productivity."
"Code assistant tools can easily generate wrong or misguided output that can cause problems to snowball."
"It is important when using code assistant agents, or any agent for that matter, to verify their output with human oversight and ultimately to ensure they are tools that work for you rather than the other way around."
"If you're looking for the most useful responses possible from a chatbot agent, then it's best to be as explicit as possible with your words."
"The same properties that make AI agents useful, they can also be dangerous. Unsuspecting users can and have provided agents confidential company and personal information that can compromise critical company intellectual property or infrastructure."
"We are flying blind into what has been called the fourth industrial revolution."
"The question moving forward with AI agents isn't whether they will be used but how they will be used."
"If you are at all interested after having read this report, there is no better time than now to get started."
AUTHORS EXPERTISE
Benjamin Labaschin is principal machine learning engineer at Workhelix, where he develops generative AI and large language models and applies them to economic problems. He previously led machine learning initiatives at companies including Hopper, XPO Logistics, and Blackstone. His career spans both applied ML engineering and the strategic questions of how AI tools fit into business workflows.
"What Are AI Agents?" was published by O'Reilly Media in November 2023 as part of their report series. O'Reilly is a leading publisher of technology and computing books, known for their animal cover series and their focus on practical, developer-oriented content. The report reflects O'Reilly's editorial approach: concise, technically grounded, and written for practitioners who need to get up to speed quickly.
Labaschin's writing style is deliberately accessible. He uses analogies (the car and engine metaphor), fictional scenarios (Very Fake Company, Inc.), and concrete tool demonstrations to make AI agent concepts approachable for non-specialists. The report was written with text generated by OpenAI's ChatGPT-4, a detail Labaschin transparently discloses in the footnotes, making the book itself an example of AI agent-assisted writing.
RESOURCES
What Are AI Agents? by Benjamin Labaschin: https://www.oreilly.com/library/view/what-are-ai/9781098159719/
Hands-On Large Language Models by Jay Alammar and Maarten Grootendorst (O'Reilly): https://www.oreilly.com/library/view/hands-on-large/9781098150945/
Chainlit documentation: https://docs.chainlit.io/
HuggingFace Chat: https://huggingface.co/chat/
OpenAI ChatGPT: https://chat.openai.com/
NEXT STEPS
Identify one workflow bottleneck: Pick a single repetitive task in your work that involves documents, questions, or code, and commit to testing an AI agent on it this week.
Write your first system prompt: Draft explicit instructions for an agent that define its role, constraints, and output format. Test it, refine it, and save the version that works.
Build a verification habit: For every agent output you act on, document what you checked and what you corrected. Make human oversight a reflex, not an afterthought.
Audit your data exposure: Review what information you have shared with public AI agents. Remove anything confidential from your workflow and move to a private solution if needed.
Track costs and value: For one month, log what you spend on AI agents and what time you save. Use the data to decide whether to scale up, switch tools, or cut back.
Watch the regulatory horizon: Set a monthly reminder to check AI legislation news in your jurisdiction. Anticipate rules before they affect your operations.
Comments