The super AI agents from Anthropic or H are the stuff of dreams, but without understanding end-to-end workflows, quantitative algorithms and optimal integration into the IS, the promise quickly turns into a gimmick.
In recent days, LLM champions, notably Anthropic, or dedicated companies such as H, are increasing the number of announcements about their “super agents”.
They can perform tasks for you on the Internet using your browser or enter them into the company’s information system.
Yes… But…
There are a few additional dimensions and limitations to consider.
First of all, carrying out the task of a “persona” in a company requires understanding a workflow across several systems with business optimization logic. If this optimization is a little complex, for example an optimized intervention team deployment schedule, a transport logistics plan or a pricing decision for hotel yield management… it requires technological bricks complementary to large models, such as Python code or optimization solvers.
This highlights two dimensions: the business analysis of the workflow (which must be iterative and pragmatic, and cannot be effectively captured by specifications written upstream) and the use of technological bricks other than those of pure generative models. To put it simply, models deal with letters, but in life you have to deal with numbers, optimizations and processes.
Second limit, that of interaction with the company’s information systems.
We will not return to the dimensions of cybersecurity and large-scale operational resilience which are obviously prerequisites.
Simple integration into an information system to allow the AI agent to imitate what a human would do requires understanding the user interfaces of the systems (SAP, Salesforce, Outlook, Servicenow, Documentum, etc.) and reacting to the context (screen change, particular case of the process, data to be entered that is unsuitable for the format expected by the application). These problems are well known to those who have tried to deploy Robotic Process Automation (RPA) for years.
On this point, the current model’s gait is clumsy like that of a baby taking its first steps while staggering. They explore each screen before each action which is not only very time consuming but very costly in tokens.
Other, more elegant approaches exist in constructed and orchestrated agentic systems. We can, for example, use models to explore and define the process and its different scenarios then treat them in a deterministic manner as an RPA would, while maintaining agentic monitoring to allow us to regain control in the event of a problem and reanalyze the situation (in self-healing or self-healing mode).
Finally, the strategy for moving into production and monitoring the performance and relevance of agents must be precisely calibrated between speed of materialization of gains and risk for the company… All this must, finally, be supported with regard to employees, authorities, partners, without forgetting customers!
Conclusion, these new agents provide definitely useful building blocks but although they constitute an ingredient in the recipe, they do not, on their own, make it possible to prepare and serve dinner!
