Generative AI is moving at warp speed, and after spending a good chunk of 2024 playing around with different models and tools, as we are welcoming the new year, I’m bursting with excitement for what 2025 will bring. We have seen the usage of Generative AI evolving from chatting with a model (call it a Chatbot) to GPTs/Copilots (Chatbots specialized in a specific skill) to AI Agents (AI which does physical work) in the year 2024.
I’ve been particularly focused on a few key areas, but not limited to them, that I think will change how we build and use software in the new year and beyond. Here’s what’s got me buzzing:
1. Code that Writes and Executes Itself : Code Interpreter
I’ve been amazed to see how well LLMs can generate code. Imagine asking an AI to solve a specific problem, and it instantly spits out the code on the fly, executes it, and generates the result you are looking for! That’s the power of dynamic code generation. A simple example: ask ChatGPT to calculate a home loan EMI with your desired inputs, and you will see it will write a Python code snippet, feed the input parameters supplied by you to calculate the precise EMI.
What’s even cooler is making sure this code runs safely. Tools like Azure’s Assistant API, which uses OpenAI’s Code Interpreter, and secure execution environments like Azure’s container app dynamic sessions are crucial here. In 2025, I want to really dig into making this work for complex problems and, of course, making sure everything is super secure and explainable.
2. Turning Software into Smart Agents: The Rise of AI-Powered Assistants
Think about all the software we use every day – APIs, functions, and so on. What if these could become smart agents, working together to achieve complex goals? That’s what I envision for 2025. I feel like existing software assets will naturally progress toward intelligent apps or agents. Imagine an agent that needs to book a flight and a hotel. It could automatically discover and use existing APIs for flight booking and hotel reservations, combining them seamlessly. To make this happen, we need to make sure our code is “agent-friendly” – that means clear documentation, ideally using standards like OpenAPI. This lets agents understand what each piece of software does and how to use it by looking into its well-written specifications. So far, developers used to write API documentations for another technical person (read: developer) to understand; now it has to be understood by an agentic system.
3. The Power of Teamwork: Multi-Modal-Model Multi-Agent Systems
One agent is good, but a whole team of specialized agents is even better! I’m really excited about the idea of multi-agent systems, where different agents, each powered by the best LLM or SLM for its task, work together to solve complex problems. Imagine an agent specializing in image analysis working with another agent focused on natural language processing. And to make sure everything works smoothly, we’ll need “critique agents” to review the work of other agents and provide constructive feedback, just like a good editor! And the feedback gets implemented to improve the quality of the outcome. Microsoft’s AutoGen Studio is a great place to start building multi-modal-model multi-agent systems; I’m looking forward to exploring it and other similar platforms.
4. RPA Gets a Brain: The Future of Automation
Traditional RPA (Robotic Process Automation) is pretty rigid – it can only follow pre-defined steps. But with Generative AI, I think RPA is about to get a serious upgrade. Tools like Microsoft’s Magnetic One are showing us what’s possible: automation that can actually interact with computer screens, browsers, and even write code on the fly. And if paired with reasoning models like OpenAI’s O1 or O3, it will self-reflect on each step of the task and potentially rectify errors. This means we can automate much more complex processes that involve interacting with different applications and systems with self-healing intermediate failures in between. I want to watch this field and see how it evolves to fully or semi-autonomous systems, keeping humans in the loop.
5. Business Logic that Updates Itself: Adapting to Change on the Fly
Changing business rules often means rewriting code, which is time-consuming and error-prone. What if we could automate this too? I see the potential of using RAG and dynamic code generation to translate plain English business rules directly into code and execute them. This would allow systems to adapt to changing requirements much more quickly. There are challenges, of course, like ensuring accuracy and security, scalability, and many more, but the potential payoff is huge. I’m really excited about its potential.
6. AI for Everyone: Democratizing Generative AI Development
Ultimately, I believe any developer should be able to build amazing Generative AI applications, regardless of their AI/ML background. I strongly believe if a developer has a good grasp of programming and cloud knowledge, they can start building Generative AI apps right away for most use cases. I’m not discounting the fact that complex use cases where model fine-tuning or building Small Language Models are needed will definitely require advanced AI/ML data science developers. That’s why I’m excited and want to advocate for the democratization of Generative AI development for all developers, making it easier for every developer to build intelligent apps and harness its power.
2025 is shaping up to be an incredible year for Generative AI, and I can’t wait to see these exciting developments come to life. Happy New Year.
The views and opinions expressed in this article are solely my own and do not necessarily reflect the views or opinions of my employer.
