Agentic AI: Thinking inside the box
Foreword by Warren Pearson, CTO of Insig AI:
Here at Insig, it’s part of our job to keep up to date with the fast-moving world of AI. With Agentic AI being hailed as the latest Next Big Thing, our summer interns were tasked with extensively researching it, to present an educational session to the team on its uses, its applicability to our own processes, and its overall maturity. Agentic AI is clearly impressive, and its potential is vast. However, its limitations, which are explored in this article, are equally important. At Insig, we are now actively evaluating selective, controlled use cases for Agentic AI, allowing us to realise early benefits while staying conscious of current limitations.
Foreword by Warren Pearson
CTO of INSIG AI
Just as we were getting our heads around generative AI and its constant updates, the next iteration has arrived: Agentic AI. If you work in the tech industry, chances are high you’ve encountered the term.
But what is Agentic AI? And is it quite as revolutionary as its marketing suggests?
What is it?
Agentic AI refers to systems that can autonomously plan, act, and adapt towards an end goal with very little human input. Essentially, it is AI that not only answers a question, but builds a strategy, breaks it into sub-tasks, gathers the data, executes code, and then checks whether it performed to a good enough standard, before refining the process once more, if it thinks it can improve… all while you’re busy making your morning coffee!
Is it that good?
On the surface? Absolutely.
Agentic AI has the ability to complete in five minutes what would take a human analyst ten days. In sectors like finance, where time is invaluable, this efficiency could be game-changing. However, speed alone should not be mistaken for depth or intelligence.
The illusory genius
If you have ever asked ChatGPT to explain something you are already an expert in, the flaws show themselves. It may produce a decent summary: coherent, passable, and occasionally eloquent. However, it rarely produces an answer that challenges your thinking, suggesting its capacity for original insight is close to zero.
Agentic AI can be understood as more advanced than Generative AI, since it can act autonomously based on an initial prompt; completing multiple loops of planning, acting, observation and iteration. However, it still does not understand, in the human sense. It lacks the ability to think critically: thinking ‘outside the box’, the way a human can. As convincing as it may appear, it remains a probabilistic model that can exclusively operate within the box it was trained in.
Humans vs. Agents
Agentic AI isn’t going to singlehandedly invent breakthrough technologies or produce original works of art. Without intuition or the ability to emotionally resonate as humans can, Agentic AI cannot understand nuance and context, apply emotional intelligence, generate novel ideas, critically think, and learn through lived experiences.
It can, however, assist and accelerate these processes. But importantly, it does not possess the capacity to replicate them – its role being one of augmentation, not replacement. The advent of agentic AI will accelerate the automation of routine and data-heavy tasks, but fears around AI replacing human thinking remain largely overstated – for now.
A significant part of the research project involved testing out different agentic AI systems. MiniMax stood out thanks to its impressive range of functionalities. It is a global AGI company producing multimodal AI agents, able to generate presentations, images, videos, music, reports and more.
We tested MiniMax’s capabilities for both presentation and website creation. At surface level, the Agentic AI powering these functions was truly impressive. The outputs produced were generally aligned with the prompts and were visually engaging.
However, limitations quickly became apparent – with lack of editability and flexibility, and the failure of certain interactive features on the website it generated. Ultimately, the digital assets it generated were unusable. However, given how fast these platforms are evolving, these tools will become increasingly useful, as limitations are resolved.
Inside the massive box
So what makes Agentic AI seem so… powerful?
Because the framework, or box, it operates within – while still a box – is unimaginably vast. These systems can access data across industries, formats and languages, with a speed and scale human brains cannot begin to match. This alone can make Agentic AI seem omniscient.
But, as we’ve seen, Agentic AI is neither magic nor omniscient. It draws from what is already known, without capacity for imagination or novelty. It’s powered by a vast store of meticulously indexed data that can deliver answers quickly, but it remains confined to its training and the limits of its box.
To conclude
Agentic AI doesn’t replace human intelligence – it can however extend it. Used as a tool, it could provide the basis for automating tedious and time-consuming tasks. By using it for this purpose, it frees up time and mental bandwidth for humans to engage with deeper thinking, creativity, and innovation.
So yes, Agentic AI is impressive. But if you’re an expert in your field, it’s unlikely it will out-think you. You can, however, expect it to support you. If you learn how to harness what is inside the box, imagine what that could unlock behind it. At its core, it’s no match for the creative mind – it’s just remarkably good code.