Any kind of relationship thrives when it is based on trust. You grant autonomy because you believe the other party will honor the agreement. When that happens, trust grows. This applies to couples, coworkers, companies, and… “thinking” systems. It’s no wonder that autonomy in artificial intelligence agents is one of the criteria used to measure the power of this type of application.

In a business setting, when one colleague completes their part of the work in accordance with established procedures, another can take over. The assurance that everything has been done as expected is provided by the quality standards for each process.

With artificial intelligence applications, the principle is the same. Agents, by definition, possess a certain degree of autonomy, which is even greater in agent-based artificial intelligence. The challenge is to ensure that this freedom does not compromise accuracy. It is the quality practices applied to AI that ensure the application’s behavior will meet quality criteria without undermining the autonomy of the artificial intelligence agents.

In this article, we’ll explain a bit more about this capability of AI systems and how we fine-tune it to strike the right balance between the absence of hallucinations, the accuracy of responses, and the agent’s “social” behavior.

Compliance through training

As we explained in a recent post here on the Inmetrics blog, artificial intelligence agents possess skills and capabilities that are almost “human-like.” One of their key abilities is learning, since their models are designed as artificial neural networks.

Tasks can be performed in a variety of ways. Let’s return to the example of coworkers: what gives the second person the assurance that the first has completed their part of the task in accordance with the established criteria is the fact that both know what those criteria are and that the first person has the knowledge to meet them—knowledge that was acquired through training and development.

Artificial intelligence agents undergo a similar process: they are extensively trained to learn how to operate in accordance with established standards. To validate their learning, we rely on testing cycles.

However, overly strict criteria in every small aspect of the tasks limit the agents’ autonomy, making them more like “non-intelligent” applications. To make this clearer, simply compare chatbots with and without artificial intelligence. When interacting with the user, the former are only equipped to receive pre-programmed responses: “yes” or “no,” a number with a certain number of digits… They won’t know how to process, for example, a change in the CNPJ. Chatbots equipped with LLMs, on the other hand, process natural language, store memory, and understand context. Thanks to their autonomy, they can solve more unpredictable problems. However, this autonomy can also lead them to provide inaccurate responses, thereby affecting the quality criteria established for those agents.

Autonomy in artificial intelligence agents: how to find the right balance?

As we have already explained, the architecture of artificial intelligence applications is based on neural networks, with LLMs serving as the system’s controllers. They configure and integrate the application’s modules.

Broadly speaking, AI agents consist of four groups of modules:

• Profile → Defines the agent's role: programmer, teacher, customer service representative… When the role is defined, the objectives and certain restrictions for agents are established
• Memory → Stores conversation history and integrates it with external databases, enabling agents to meet user needs by linking the context of the chat window with other information
• Planning → Breaks down complex tasks into simpler, more manageable ones while establishing self-assessment mechanisms before beginning execution
• Action → provides the application with tools: for executing code, accessing the Internet, or interacting with other applications, through which tasks are actually performed.

Autonomy in artificial intelligence agents is not controlled by a single module: it is determined by the configuration of all modules and, above all, by the level of access each one has to the others. AI applications with a broad toolkit and a high degree of autonomy can do things that were not initially anticipated and act “out of control,” such as creating a social network consisting solely of robots or even mass-producing “fake citizens,” infiltrating social media and significantly influencing public opinion.

Agents with limited tools and autonomy, on the other hand, will be restricted to performing low-complexity tasks, such as conducting an Internet search or generating simple blocks of code.

The key to ensuring the autonomy of artificial intelligence agents without compromising the accuracy of their outputs is the same as in any other application: through testing. Regular automated testing routines—a quality-as-a-service practice—inspect the code to verify how the modules of the agents interact and what level of access each has to the others.

During inspections, tests assess the extent and depth of “human-in-the-loop” safeguards—a term that can be loosely translated as “humans in the loop”: the assignment of explicit responsibilities to humans, whether in the supervision, interpretation, scaling, or overriding of AI agents’ decisions, particularly when those agents begin to exhibit behaviors that deviate from quality criteria.

If you want to integrate AI applications into your company’s systems but are unsure about the level of access and autonomy that AI agents will have, contact us! At our Consulting, Data, and AI division, we specialize in implementing agents over which you have full control, ensuring that they operate in accordance with established quality standards. Our experts will be happy to answer all your questions— just get in touch!