The AI industry is accelerating rapidly, and this is evident in the introduction and application of AI agents. A few years back, AI was just an LLM wrapper, but now, with options like Operator by OpenAI, it is much broader.
At MLDS 2025, India’s largest GenAI summit for developers, organised by AIM, Siddhant Goswami, co-founder of 100xEngineers, an AI lab specialising in Generative AI, explained how AI agents have evolved and changed over the past few years.
Goswami noted that over the past two years, ever since he started developing AI Agents, their definition has changed significantly each year and each quarter. He referenced “God In A Box,” which many had initially perceived as an AI agent, but it merely was an LLM wrapper. He emphasised that perspectives are shifting and that the entire industry is advancing rapidly.
Not just Goswami, but many Indian founders, in general, love AI agents.
The Phases of AI Agent Evolution
The journey of AI agents can be divided into distinct phases, each introducing a new level of sophistication and functionality. Initially, everyone was clueless about why they existed, but they are now building meaningful solutions based on this knowledge.
Goswami outlines five phases of AI development. The first phase began with large language models (LLMs) like ChatGPT in 2022, which could process text inputs but lacked advanced reasoning. The second phase improved context windows, allowing users to input extensive texts, leading to the practical use of corporate databases.
The third phase introduced retrieval-augmented generation, enabling LLMs to access external knowledge. The fourth phase brought multimodal capabilities, allowing LLMs to process images, voice, and video. In the fifth phase, LLMs utilise memory layers, enhancing their ability to remember and adapt responses.
Building an AI Agent as an Artificial Being
Summing up the phases of building an AI agent, Goswami told AIM, “We are kind of recreating a being, an artificial being. We first gave them hands and legs, then we taught them to learn the language, and then we gave them memories so they could remember things. I feel that’s the evolution of AGI itself.”
With this type of advancement in AI agents, he believes we can expect to achieve AGI sooner.
Especially when Indian founders, in general, love AI agents.
AI Agents Need Human Intervention
At MLDS, Goswami mentioned that AI agents can hallucinate and go off the rails. So, a human-in-the-loop model is needed to monitor the operations.
“Even if you build the best technology right now, without having a human in the loop and giving feedback, I don’t think we would be able to achieve the level of intelligence we are expecting from this.”
Despite the exciting progress, Goswami warned that not everything requires an AI agent. Many businesses rush into agent-based automation without understanding the trade-offs.
Most AI problems can be solved with a simple LLM and a RAG setup. Companies implementing AI solutions should prioritise simplicity. He suggested not to overcomplicate things and only introduce agents if necessary. If a task requires multiple steps, reasoning, and external tools, that is when we should consider using an AI agent.
What’s Next for AI Agents?
It looks like AI agents will become increasingly autonomous, with longer memory retention and real-world action-taking abilities.
Research and Market’s report on ‘AI Agents Market Analysis projects the market for AI agents to grow from $5.1 billion in 2024 to $47.1 billion in 2030, with a CAGR of 44.8% during 2024-2030.
For now, with human observation added, the definition of AI agents should continue to evolve. It’s exciting to see what the future holds for them.
