Skip to content 
2025 has been heralded as the year of AI agents. Interestingly, while the impact of their wide adoption is underscored, their capabilities are often misunderstood.
To understand the nuanced reality of using AI agents at scale, we need to explore where we are today with language models and agents.
This content is also available in audio format. Listen here:
In November 2022, ChatGPT was released to the public to gain millions of users and reshape the industry almost overnight. What began as an experiment for developers at OpenAI ignited a global AI revolution.
Since then, we have seen a wave of new models and the integration of generative AI in countless products. These models are not just powering user-facing tools like chatbots; they are working behind the scenes in ways you might not even notice. For example, hyper-personalisation features now craft a massive portion of the text, images, and videos you see online. With little human intervention, these AI systems design content that evokes emotions, grabs attention, and helps brands connect more effectively with their audiences. The data that powers them is the reason an ad can feel like a sign and your reel feed is infinitely engrossing.
At the heart of these advancements are large language models (LLMs). While they enable immense and rapid innovation, they are not without limitations either. LLMs’ knowledge often lags behind, and they are cut off from interacting with the real world.
To put in simply, LLMs only understand and produce text. While their assistive power is transformative, helping many of us work smarter and faster, they fall short of what is truly possible.
This is where AI agents come in to power the next leap forward. To understand how AI agents work, consider how this article was created. It definitely did not emerge fully formed in one go, but it is rather the product of the following steps:
- Concept: The ideas behind this essay have been coming together over weeks. I have been thinking about them for a while, and one day during a workout, the structure of the article finally clicked;
- Research: I spent a few hours exploring different perspectives to deepen my own understanding;
- Coding: While I didn’t write code for this piece, my experience building LLM applications gave me insights that inform the explanation you read here;
- Outline: I drafted a bulleted list of the key sections;
- Draft: I fleshed out the first version of this article;
- Edit: I refined the draft to make it more clear and concise;
- Revision: A content marketer revised the article to make it more impactful and prepared it for publication.
Now, imagine asking an LLM to create an essay in one step. It would be like asking a person to sit down and write something with zero research, brainstorming, or editing. In other words, a task likely doomed to failure. Here, AI agents make all the difference.
Agents are designed to solve two massive challenges of LLMs. First, they change how we interact with these models, helping to overcome the zero-shot generation problem by enabling multistep processes. Instead of treating tasks as one-off requests, agents orchestrate workflows that handle each step logically.
Second, they address the connectivity gap. LLMs are confined to their internal world: they cannot access live data, store memory, or interact with real-world systems. AI agents bridge this gap, unlocking the ability to retrieve information, remember past interactions, and even act on behalf of users.
Now that we know what the challenges are, let’s dive deeper into the technological breakthroughs that enable AI agents to tackle them.
Zero-shot generation
The first challenge is solving the problem of zero-shot generation. Currently, several frameworks allow developers to orchestrate multiple calls to an LLM, forming the foundation for agentic workflows. These workflows are designed to handle tasks in a systematic sequence: analysing, planning, generating, reviewing, and finally acting. The first four steps can repeat as many times as needed to refine the outcome.
Each framework has its own approach to reasoning, and most rely on one or more of the following:
- Chain-of-Thought (CoT): This approach breaks down complex problems into smaller, logical steps, making it easier for the agent to work through challenges step by step;
- ReAct (Reasoning and Acting): Reasoning and acting are interwoven, allowing the agent to both think and take actions dynamically as it processes tasks;
- Tree-of-Thoughts (ToT): Expanding on CoT, this method enables the agent to explore multiple possibilities, evaluate them, and backtrack when necessary, much like strategising through a decision tree.
These reasoning methods rely on advances in prompt engineering, which help tailor how LLMs respond to tasks. By combining reasoning and action capabilities, these frameworks facilitate the step of acting for AI agents.
A number of tools on the market offer varying levels of orchestration, abstraction, and model interoperability. Here are a few noteworthy examples:
LangChain / LangGraph
An open-source library offering prebuilt workflows for common tasks such as Q&A over a database, chatbots, query analysis, summarisation, autonomous agents, and more.
AutoGen
Microsoft’s open-source framework takes a high-level approach to multiagent conversation. AutoGen allows for customisable agents that communicate with one another to solve tasks. Its flexibility makes it suitable for building a wide range of LLM applications.
LlamaIndex
This tool specialises in context-augmented LLM applications, offering connectors for over 160 data sources and formats. It also includes modules to evaluate application performance, making it a robust option for integrating diverse data into workflows.
Connectivity gap
The second major problem with LLMs is connectivity. LLMs, by design, lack access to real-time information, cannot recall past interactions, and cannot act in the real-digital world. To make agents effective, they need to bridge this gap. There are several technological advances designed to enable this:
Retrieval-Augmented Generation (RAG)
RAG enhances LLMs by enabling them to search external sources like web pages or knowledge bases. The retrieved data is pre-processed and integrated into the model’s context, giving it a broader understanding of the topic.
For example, instead of relying solely on its pre-trained knowledge, an LLM equipped with RAG can generate more accurate, detailed, and contextually relevant responses.
Tools
This category includes advancements that let LLMs interact with the real-digital world:
- File search: Implements RAG to retrieve information from documents;
- Structured output: A filtering mechanism within LLMs that ensures outputs adhere to a specific structure and is useful for building deterministic applications;
- Code interpreter: Allows LLMs to execute code in a sandbox environment to enable tasks like performing calculations or data analysis;
- Function calling: Perhaps the most revolutionary feature, enabling LLMs to reason and decide how to solve problems by calling external functions.
For example, imagine you ask an agent to provide the current weather. While an LLM cannot access live data, an agent equipped with function-calling capabilities can retrieve real-time weather information via an external API. This concept scales infinitely: agents could send emails, book holidays, or even react to changes like a human assistant would.
It is no surprise that OpenAI CEO Sam Altman predicts we are going to see the first agents “join the workforce” in 2025. These agents would be able to perform a wide range of tasks, dynamically adapting and responding to real-world needs.
Think about how many tasks AI agents could automate. While many of these tasks can already be handled through APIs, not everything fits neatly into that category. For those tasks, we are now seeing the rise of computer-use modes, where agents can observe your screen and control your mouse and keyboard, quite literally replacing you in some activities.
Business impact in the short and the long run
One significant barrier to organisations adopting AI agents is the absence of a Data & AI: Fast and Slow (DAFS) platform. Such a platform, as described in our eponymous book, captures all data of all types and enables it to flow freely between systems, people, specialised AI agents, and niche machine learning models. This seamless flow allows for real-time decisions and actions that are inherently data-driven.
So far, many organisations have managed to avoid fully modernising their data and AI capabilities. They have been able to deploy niche machine learning models that rely on smaller datasets and automate predefined processes. While a DAFS platform would have made these efforts faster to deploy and experiment with, it had not been essential for one or two isolated use cases.
Over the next 5–10 years, however, this will change dramatically. Foundational DAFS platforms that integrate structured and unstructured data at speed while leveraging the wisdom of historical, slow data will become critical. These platforms will underpin the next generation of AI workers, enabling organisations to harness real-time insights and actions on a much larger scale.
Even today, more than 80% of enterprise business leaders say that data integration is critical to their success. Yet, many enterprises are held back by a proliferation of legacy systems, which make real-time data integration and the use of AI agents nearly impossible.
Legacy data integration platforms like Informatica PowerCentre and Talend come with significant drawbacks that limit their effectiveness in an AI-driven world. To name just a few:
Lack of documentation and skilled expertise
These systems often lack adequate documentation, which makes modernising or upgrading the platform cumbersome and time-consuming. Adding to the challenge, there is a shrinking pool of experts skilled in these platforms. Even if companies find the right talent, retaining them can be costly.
Reduced business efficiency
Legacy platforms can hinder productivity for several reasons:
- They are built on outdated frameworks that struggle to integrate with modern systems;
- They lack connectors to newer SaaS technologies, databases, and IoT devices;
- They cannot handle exponential data growth or process modern queries effectively;
- They are incompatible with cloud migration and multi-cluster data processing.
DataOps challenges
Enterprises relying on legacy systems tend to have rigid, siloed operations, making it difficult to implement modern DataOps practices. This limits the use of CI/CD pipelines and experimentation.
Security risks
Legacy platforms are highly vulnerable to security threats. As these systems reach their end of life, vendors stop offering support or updates, leaving them exposed. Even when updates are available, a lack of skilled experts often delays their implementation, further compromising security over time.
Risk-free modernisation
At Infinite Lambda, we see a growing shift away from legacy data integration systems like Informatica and Talend. To support this transition, we have developed Flowline, a risk-free data modernisation solution.
Our approach not only migrates your data and refactors your code but also builds a foundational platform designed for the future of AI agents and whatever comes next in this AI revolution. By adopting the modern data stack, organisations can overcome the challenges of legacy systems and unlock the full potential of data and AI in their operations.
The next wave of technological progress depends on leaving behind outdated systems and embracing platforms that enable flexibility, scalability, and real-time decision-making. With a DAFS platform in place, organisations will not simply keep up with the AI revolution but achieve meaningful, sustainable growth in modern business realities.
What’s the verdict?
So, is an AI agent lurking to take over your job? The answer depends on the nature of your work. Jobs focused on performing predefined functions, rather than creating or defining new ones, are at the greatest risk. Data entry, scheduling, and other repetitive tasks are already being taken over by AI agents, virtual assistants, and automated workflows. Industries like insurance, banking, and customer service are quickly adopting these technologies to streamline back-office processes.
This sounds like doom and gloom for many, but let’s not get ahead of ourselves. To understand the context better, I suggest taking a step back to look at how previous technological revolutions unfolded.
The Industrial Revolution
The Industrial Revolution was one of the earliest examples of widespread technological disruption. Mechanised manufacturing, powered by innovations like the steam engine and textile machinery, drastically reduced the need for manual labour in certain areas. This shift caused significant anxiety among artisans and workers, leading to movements like the Luddites, who protested against the new technology.
However, while jobs in traditional handicrafts declined, the revolution sparked entirely new industries, from railroads to mass production. These advancements created far more jobs than were lost, reshaping economies and society in the process.
The rise of computers
A more recent example is the computer revolution of the late 20th century. Early fears warned that computers would replace clerks, accountants, and even engineers. And yes, some jobs were automated. But computers also gave birth to entirely new industries — software development, IT management, digital marketing, and beyond — whose very existence we are now perfectly used to. Entire sectors of the economy, like finance and healthcare, now depend on professionals who use and innovate with computers.
In both the Industrial Revolution and the computer revolution, fear of job loss was real, but so was the potential for job creation. History shows that while some roles are made obsolete, new opportunities emerge, ones that align with the technology of the time. For example, while traditional factory jobs declined, roles for engineers, machine operators, and data analysts grew.
The rise of AI agents will likely follow a similar pattern. Yes, repetitive, routine jobs are at risk of being replaced. But history tells us that disruption comes hand-in-hand with creation.
New opportunities will emerge in areas that require uniquely human qualities like creativity, emotional intelligence, and advanced problem-solving. For example, as AI handles routine tasks, demand may grow for people who can design, oversee, and improve AI systems or use them in innovative ways.
The lesson from past revolutions is clear: technology disrupts, but it also creates. The challenge for us is to adapt, learn new skills, and position ourselves for the opportunities that AI will bring.
Learn more about Infinite Lambda's legacy modernisation work.
