Let’s be honest. Asking a general-purpose generative AI like ChatGPT, Grok or Gemini to solve a complex business problem is a fascinating experience.
It’s like asking a Golden Age poet to do your tax return. The result will be impressive, but it will almost certainly guarantee you a very serious talk with the Inland Revenue.
We have been seduced by the promise of universal genius: a single AI capable of analyzing markets, writing emails, debugging code and, if it has time to spare, composing a ballad about asset depreciation.
The problem is that this genius, when confronted with the ambiguity of the real world, often suffers from what we might call “excesses of creativity” or hallucinations.
For example, if you ask an AI to analyze sales data, draw conclusions and send an email report, it will deliver a financial analysis with the prose of a novelist and the conviction of a politician, but with the precision of an octopus with a scalpel.
Decomposing complexity with an agentic system
In the face of this multitasking chaos, an elegantly logical and simple solution emerges: the distribution of tasks in agentic systems (also known as agentic AI or agentic flows).
These systems aim to mitigate the known problems of hallucination and inaccuracy of monolithic AIs, and do so by reducing the context, knowledge and tools available through agents: Specialized, task-oriented AIs.
“The agentic AI strategy is to stop searching for answers with the universal genius and instead assemble a team of specialists. A bit duller, yes, but infinitely more useful.”
Marc Jordana, Digital Lead of Softeng.
How agentic AI works
There are several models of agentic systems, but to explain the concept we will base ourselves on the most common, which is the hierarchical model.
In this model, you have an Orchestrator Agent, who is basically a project manager who doesn’t spend the day asking how things are going; he simply understands the complex problem and breaks it down into tasks for his team of specialists and coordinates the work of this team.
Let’s go back to the previous example, now in prompt format: “Analyze last quarter’s sales, identify the three worst performing products in the North and draft a mailing to the sales team with recommendations”.
A know-it-all AI could get quite creative with this prompt. The agentic system, on the other hand, would do the following:
- The Orchestrating Agent receives the order and decides on a plan of action.
- Awaken the “Figures” agent from its lethargy. This agent lives in a world of tables, cells and SQL queries. It doesn’t understand irony or sentiment, but it finds you a needle in a haystack of data in nanoseconds. It receives the request in natural language, converts it to a structured data query and returns the results flawlessly.
- With the data, the Orchestrator goes to the “Strategist” Agent. This agent wouldn’t know how to run a database query if his life depended on it, but he is able to look at the query results and generate three brilliant ideas to sell more. His specialty is “what to do with this data,” not “where the numbers are coming from.”
- Finally, the Orchestrator passes the findings on to the “Golden Piquito” Agent. A communication expert who can write an email so motivating and clear that the sales team will go out and sell even the staplers in the office.
Everyone has done their part. Without drama, in a coordinated manner, focusing on each task and without interference that can generate hallucinations.
The power of specialist agents is “not knowing everything”.
The magic of this system is selective ignorance. Agent “Figures” does not need communication skills and Agent “Goldilocks” does not care about the complexities of the database.
They communicate as little as possible and only about their particular specialties. A revolutionary concept in a world obsessed with everyone having to form an opinion on absolutely everything.
This compartmentalized approach drastically reduces the possibility of destructive hallucinations due to the ambiguity of the information and the saturation of the single model.
“AI works best when its universe and context are well defined. By giving each agent a small realm to rule, you ensure that its decisions are more reliable and predictable.”
Marc Jordana, Digital Lead of Softeng.
This is the difference between using AI as a spectacular smoke and mirrors game and applying it as a business tool that actually automates complex processes.
What do you prefer in your team, an incredibly creative genius, but with hallucinations, or a team of efficient specialists?