Agentic AI: developing the skills and systems for success

How can we manage an AI designed to build and execute its own plans, with even less human input? Agentic AI is developing rapidly, raising new questions about ethical frameworks, deployment and what skills the future workforce will need.

How would you define agentic AI and what sets it apart from traditional AI models?

The cynic in me will say that the difference is it's more expensive and likely to do more damage! But, if we take this more seriously… let's look at what the current state is. We now have AI agents, as in an AI system that can do things for you, and this has only been the case for around a year.

Currently, people are using AI agents for workflow automation. With workflows you have predefined steps, a set of tools you're going to use, and some code that's going to orchestrate some stuff. It's all very predictable and it's going to do things essentially step by step through that particular workflow for you. You know what's going to happen, when it's going to happen, and what the outcome will be — you've just got some AI agents in there to automate some of the tasks. 
What people are now talking about with agentic AI is that the AI itself is going to decide what tools it wants to use and what steps to take, and then put together a plan and execute that plan. It will assess, and then possibly iterate and change the plan, and do things differently if things aren't working. There is an enormous amount of autonomy going on with these AI agents. That's the theory. In the real world, very few people are actually making that work, and very few companies are succeeding with it. 

Readers may well be familiar with the Towards Data Science blog, which I absolutely love, and you get some real insights there as to what's really going on. Recently they had a lovely blog by an AI developer which discussed how to build scalable AI and agentic AI systems. What I quite liked was the description of what the difference is: ‘An AI workflow is your reliable friend who shows up on time. An agentic AI is the smart kid who sometimes goes wrong.’ Increasingly this actually seems to be the case, according to the large scale studies that have been carried out by the likes of Salesforce and Carnegie Mellon University. They've looked at how agentic AI systems work on real world problems and have discovered that, if you're lucky, you can get them to complete simple tasks 60% of the time. But if you're asking them to do multi-stage tasks, which of course is the whole point of agentic AI, the success rate reduces significantly. So, while these systems sound like a perfect solution, very few people are getting them to work in practice.

What skills do you believe are required to successfully adopt agentic AI, and why are many businesses struggling with successful AI integration?

The best descriptions of the kind of roles and skills that you would need to integrate generative AI across your business in a way that's going to result in increased profits is from McKinsey. They cover things like, for example, people who can do data ops, people who can manage data pipelines and make sure those are always available, and that you've got the right kind of data for an AI system to ingest. You need people who can manage the reliability and availability and performance of software in a way that works for AI systems. You've got to have the right kind of DevOps engineers for this kind of transformation. You've got to have people who can understand cloud architecture. You've got to have people who know how to write efficient, scalable, modular and maintainable software architectures, as well as having things like data scientists and data engineers.

Yet, how many firms in the mainstream economy have all of these people? The answer is, very few. So, it's not too surprising when you look at that incredible range of people that you need to fulfil those roles that you can see why most companies aren't trying to do huge scale transformation of their IT infrastructure and their business models that would enable them to successfully use generative AI. What they're actually doing is looking at small use cases where maybe they're summarising data, they're doing some code completion, they’re doing a bit of data analysis… it's nothing like the kind of scale of project where you would need an agentic AI solution.

At a time when agentic AI becomes more autonomous, how can we ensure that such systems align with human values and remain under meaningful oversight?

It's going to be difficult, because if you look at the way AI systems are designed, they are intrinsically unreliable right from the get-go.

The International AI safety Report 2025 states ‘there has been progress in training general purpose AI models to function more safely, but no current method can reliably prevent even overtly unsafe actions’. That's the opinion of many very reputable academics and leading people in the industry who've been looking at this across multiple countries; you can't trust them and they're not reliable. So, the only way this is going to work, at least for the foreseeable future, is if there is really strong human oversight. This means that when you are adopting these things, you've got to make sure that the kind of governance that's wrapped around this stuff has proper break points, proper exception handles, proper risk mitigation, proper contestability and proper openness and transparency so that humans understand what is going wrong when it goes wrong.

Do you believe current regulatory and ethical frameworks are adequate to handle the risks posed by increasingly autonomous AI and agentic AI systems?

Currently, I think it’s the case that they're racing to catch up. We've seen that with the EU AI Act when that was first being drafted. There was no such thing as generative AI really, and certainly nobody had even coined the phrase agentic AI. 

For you

Be part of something bigger, join BCS, The Chartered Institute for IT.

One of the reasons that regulation is going to struggle with this is that it's only possible for regulation to work well when you have really robust technical standards, governance standards, ethical standards and professional practice standards. And those standards need to be operationalised and made to work in the real world by professionals.

At the moment, we don't have the right definitions around what a professional is in the AI world. We don't have the standards around what it means for an AI system to be auditable. While those things are evolving, we can't be in a place where the regulation can be appropriate because we haven’t got the right standards and we haven't yet got the right professional practice.

Regulation is supposed to ensure that things are being done ethically, competently and that there is the right kind of accountability there. What we're missing at the moment is being able to ensure that the people who actually build, deploy, manage and run these systems can be held to account. For me, it’s essential that if people are in very responsible roles around these AI systems, they really must be registered with the profession.

What are your thoughts on AI’s potential, and where do you think it’s heading?

Despite the various challenges and barriers, it's still all extremely exciting. People do like to talk about the potential, and the potential isn’t being realised at the moment. But there are promising signs, especially in the software engineering world. There's been some good research produced recently by the people at MIT, and Berkeley and Cornell, which concluded that while there are massive problems at the moment in trying to get agentic AI to help you with code refactoring or cloud migration or software testing, there has been huge progress in how much help existing generative AI systems provide. In their research they were looking at, for example, Meta, who have some really advanced generative AI software testing technologies which are really useful 70% of the time — so, 70% of the time you can go to your generative AI testing harness, and it will work out test cases for you and help you run them. But, that still means you've got to know which of those possible test cases are useful or not useful and that can only be done by having expert software engineers reviewing those test cases.

In the future tech job market what other sort of skills will be in demand?

For the next few years, I think there will be a transition from wanting people who can churn out lots of code to wanting people who are very good at understanding code that has been written by somebody or something else, evaluating that code and identifying how and where it needs to be improved. That is a real skill that's hard to come by. I think the same is going to be true across wider business, anywhere where generative AI is producing outputs which are critical to some workflow or process or decision that's going to be made. You will need very knowledgeable people who are able to evaluate that output and determine which parts of it are genuinely useful and meaningful, and which parts of it could be wrong or misleading. And then what to do about those parts of the output that you don't like. We're actually seeing more and more prompt engineering disappearing because, increasingly, you can get the LLMs to tell you what was wrong with the prompt in the first place, and then automatically give you a better prompt.

Finally, for someone looking to break into today's tech industry with a focus on AI, what's the one key step, or one important piece of advice they should be taking on board?

If you want to get hired, learn how to make AI systems that are reliable, that are robust, that are testable, that are maintainable, that are flexible, that are extendable. Don't just make something shiny that looks nice.