Too many agents, too little sense. Business inhibits with the vision of fully autonomous AI agents
Since one agent helped in the staff, and completely different in analyzing retail trends, you had to know which one to enter and when to switch between them. This is not magic, but the troublesome logistics of the user's experience. That's why Walmart is currently taking a step back: He consolidates everything in four “superagents”. These will be separate solutions for clients, employees, engineers, sellers and suppliers.
Each suitagent is to be one gate to a number of smaller agents hidden in the background, so with one contact point instead of the tangling of applications and panels. Sparky – Superagent for customers – already works, Marty for suppliers is to debut in the coming months, and tools for employees and engineers are planned for the next year. The whole thing is fastened with Anthropic Model Context Protocol (MCP). It is worth adding that Walmart started with agents when MCP was not available and now he must rewrite older applications to speak the same language.
Suresh Kumar, Cto in Walmart, says that this “the simplification of the interface is a condition of adoption. Too many agents in too many places turned out to be simply non -intuitive.”
Check also: COMPISTING TO THE SEJM COMMISSION with Zbigniew Ziobra. “Please leave the room”
Working with AI is not that simple
The history of Walmart shows that The problem is not only “or agent can”, but “how a man is to use it”. The introduction of too many intelligent modules can complicate the organization in the same way as a poorly designed structure of departments. Instead of flat automation, we get too many decision paths, and the user who was supposed to gain support is losing time orientation in the system.
On the other hand, at the engineering level, voices of reason are getting louder. Utkarsh Kanwat, a system engineer working on several production agents of AI in development areas, devops and data, cools expectations of “autonomous wonders”. His argument begins with mathematics: If each step in the flow of the agent's work has 95 % Effectiveness (which is optimistic for today's models), it is five steps only 77 percent. successten – 59, and twenty – only 36. Even to increase the reliability of a single step to 99 percent. It only leaves about 82 percent Effectiveness at the process, which consists of twenty steps. It is not a matter of a better prompt, but a simple multiplication of probability. When we strive for production standards at 99.9 percent, the growing chains of activities without human control become statistically unsurpassed.
The canvat does not end with numbers. It shows how to design agents to make them actually. His recommendation is: Short, easy to verify work fragments, bright withdrawal points and gates with confirmations from man. In his systems, AI takes three to five steps, and each of them can be reversed or corrected before the consequences go further. This approach has a common denominator – these are boundaries and conscious inclusion of people where the risk is greatest.
Read also: Who conducts foreign policy better? Poles indicated the favorite [SONDAŻ]
The costs of AI agents can increase rapidly
This does not end with restrictions. The costs of tokens in conversation agents are growing faster than many managers want to admit. If each subsequent step has to “remember” the whole earlier contexts, the bill increases in a way that in longer interactions becomes damned. The canvat cites an example of an agent's conversation with a database, where the agent made 100 exchange of information – in this case the company had to pay from $ 50 to $ 100. for used tokens. This is not scalable for large companies. This is why The most effective agent solutions are usually quite simple. So we get a description, generate a function, we finish. Zero long memory for individual steps, zero rising costs.
The most important and often overlooked aspect is the design of tools that AI agent uses. Calling API is just the beginning. The tools must correspond to the language that the agent can effectively process in a limited context. Instead of flooding it with thousands of records, it is better for the database to return a summary of such a type of: asking, the result is 10,000. poems, here are five examples. You also need to think about partial successes, error service, relationships between operations.
This is traditional system engineering, not the magic of language models. In corporate environments, we rarely find “perfect API” to connect – there are various limits, old systems, compliance, audits. The agent will generate an SQL inquiry, but this classic software must ensure that connections, transactions, timeocas and audit logs. Ignoring these realities ends with demonstrations at conferences that do not translate into stable implementation.
A joint conclusion from both stories – Walmm and the aforementioned developer – is simple, although not necessarily convenient. AI agents will be more and more important, but not in the form of dreams of full autonomy. They are rather intelligent modules operating within strictly designated boundaries, sewn with coherent architecture and interface, with a man or a deterministic system as a quality controller. An attempt to automate everything at once leads to accumulation of errors, costs and chaos of interfaces.
An attempt to build one agent for everything without internal standards and communication methods causes the same trouble as having dozens of agents without coordination.
See also: 73-year-old Steven Seagal withdraws from businesses in Russia
AI agents. Compromises are necessary
That is why the Walmart Movement towards Superagents is an interesting compromise. The company does not give up agents, but instead of exposing everyone to the top, it hides them behind one gate of users, and underneath it standardizes the exchange of information. This is an architectural lesson, not PR. In this sense, “simplification” is really an effort for consistency and easier management, not limiting the ambitions about the implementation of artificial intelligence.
In turn, the engineering perspective reminds that If you tell the AI model to go in dozens of steps without control, the game goes against you. Mathematics does not forgive. Since you can't have 100 % Success at every step, you need to shorten the chain, design control points and separate what the agent does best, from what should remain in the hands of people or hard application logic. The best -acting AI agents in production are not talking assistants from everything but specialized toolswhich do one thing well and disappear from the way.
So if someone promises that in a moment the whole bands will be replaced by autonomous AI solutions, it is worth asking for details: how do they solve the problem of error accumulation? How do interfaces standardize? How do they deal with tokens costs and limit context? What is the restore of changes, audit, compliance with regulations? In most cases, the answer is: we don't know. Real work happens not in marketing slogans, but in the design of architecture and tools that are able to harmonize with AI on equal – and reasonable – principles. Let's stick to that.
Author: Grzegorz Kubera, Business Insider Polska journalist
