Since GenAI and LLMs took over the world, we’ve seen GenAI manifest itself mostly as chat interfaces (or API calls if you are a developer) and the biggest impact this has had on our workflows is that it has turned us humans into copy-paste machines who need to have a new skill, “prompting”. This approach to LLMs hasn’t delivered the productivity boost we expected, nor the other great promises of modern AI. Understandably, many talk about the AI bubble bursting. Agentic workflows, however, are a different beast all together.
Beyond bots: Understanding how agents work
Let’s start by examining AI agents in two ways: Through the purpose of the agentic workflows and the types of behaviours these agents have. In very simple terms, AI agents can be used for two types of purposes:
- Workflows with specific, dedicated tasks, aiming to produce a (pre-)specified outcome
- Open-ended problem-solution patterns that are based on loosely defined business problems.
This is an important perspective for understanding AI agents. They can be used in two very different ways. AI agents can be used to automate long and multilayered workflows and subsequently processes by creating dedicated agents performing specific tasks in sequence, in parallel or in loops. If you can model a workflow and define each task in the flow, you have the potential to automate the workflow using AI agents. This is reason number one for me feeling that we’ve got a bigger thing happening with AI.
Reason number two is at least equally exciting. AI agents can be set up to answer business problems without any specific pre-defined workflow or solution pattern in mind. You can assign specific roles with specific skills and instruct the agent network to collaborate in certain ways, but you don’t need to define tasks in detail upfront. You just give the network a business problem and see what happens. Then repeat this as endless simulations to see how the outcome changes and pick the best ones.
Let’s then look at the agentic behaviours in a little bit more detail. In the ‘specific workflow’ end of the spectrum, we have behaviours like responding to triggers or other user inputs (e.g. chatbots) and we have abilities to use tools like scheduling or retrieving/writing data from/to a system. Moving towards the open-ended, somewhere in the middle, we start to see behaviours like observation and reacting to events, which are useful for tasks like monitoring or operational decision-making. Finally, when reaching the open-ended exploration end of the spectrum, agents start behaving more autonomously and even creatively, where they make decisions without human prompts and they start to design and execute entirely new tasks and workflows. They could even design new agents and build autonomous teams that operate in ways they have designed themselves, without human input or prompting. Yes, this is where human oversight starts being really important. You really need to monitor the outputs, and of course, things like cost and security when humans haven’t been involved in the design of the processes.
Agents in action – what can we expect
Across industries, agents are used for a variety of use cases, and in the early days, the examples naturally tend to be in the specific workflow end of the spectrum. Within industry and manufacturing, modern predictive maintenance should include agentic tasks that not only predict issues, but also perform diagnosis and schedule service calls or other actions to prevent issues.
Within retail, we could have a network of agents performing specific tasks around upsell or next best product recommendations, combined with logistics, inventory management and dynamic pricing. These workflows have too many tasks, with too much data and context, to be given to one bot, but if we model these as workflows with specific tasks, we can let AI agents manage these in a workflow or network.
We also want to keep a close eye on new markets created or enabled by agentic workflows. Not only do they create new types of outputs and subsequently services or business models, but we also believe there will be marketplaces for agents and agentic networks. However, as mentioned earlier, it is too early to know with certainty what these new AI agent marketplaces look like and how they work, so be careful with anyone claiming they know how these markets will work.
I’m not betting all my life savings or my credibility on AI agents, but I do think these will transform, change and disrupt how we work in a fairly significant way. Let’s look at these with a healthy combination of optimism and scepticism, and do our best in implementing AI where it makes most sense.