This is guest post by Ihar Rubanau, Senior Data Scientist at Sigma Software Group. In it, he offers insights into how intelligent agents are reshaping the way we think about automation, decision-making, and collaboration between humans and machines. Drawing from his expertise in data science and applied AI, Ihar unpacks the evolving role of agents across industries and highlights their potential to drive more adaptive, context-aware, and efficient systems.
The concept of AI agents has undergone a rapid evolution. Originally, agents were simply intelligent substitutes for basic conditional logic — smart versions of switch or case statements. But things changed quickly. With the introduction of tools, agents became more than just logic handlers; they could now analyse data, search for information, and generate new content. This paved the way for a key innovation: the Model Context Protocol (MCP) — a unifying standard that allowed agents to interact with tools reliably. Some developers now compare MCP to “mega APIs,” a testament to its power and flexibility. The latest leap in this trajectory is the Agent-to-Agent (A2A) protocol, designed to enable direct communication between agents themselves — a true step toward autonomous collaboration.
Agents in action: the new computing environment
We already see real-world applications of agents in environments like Cursor and Windsurf, where agents can edit code, run files, and interact with terminals autonomously. These platforms represent the current frontier of agentic computing — hands-on, integrated, and reactive. Simultaneously, new models continue to emerge. Tools like Grok-4 are pushing the performance envelope, setting new benchmarks and hinting at what lies just ahead.
In parallel, a quiet but powerful shift is happening across the corporate landscape: many companies are slowing down hiring, likely betting that increased productivity from AI adoption will compensate for smaller teams. Yet some companies — Meta, for example — are going the other way, building elite AI teams and offering millions to attract top talent. They see what’s coming.
The Automation vs. Creativity divide
Looking ahead, what can we expect in the near term, over the next 6-12 months? Will agents truly automate every aspect of the web? Indeed, many companies are already developing AI-powered browsers where agents can access any web element and interpret user behaviour. The vision is that instead of clicking, a user would simply ask. An agent, often termed a “router,” would then analyse the query, determine the necessary tools, plan the work, execute it, and return the result.
But this raises a fundamental question:
Are we moving toward true intelligence, or just more advanced automation?
At present, most agentic systems automate existing tasks. Few go beyond what humans already do manually. They’re impressive, but largely repetitive — replicating workflows, not reimagining them. Current architectures often involve pipelines of 1–20 agents, working sequentially. Most do not dynamically interact with one another. In fact, many data scientists intentionally design them this way—to reduce unpredictability. The goal is stability and determinism, not emergent behaviour.
The shift to dynamic agentic systems
But real-world processes are rarely linear. They’re dynamic, messy, and reactive. Imagine one person updating GitHub while another updates Jira. In most current systems, information doesn’t sync automatically. A human has to build a connector or trigger a tool.
A “dynamic” approach, however, envisions “Slack” and “Jira” agents communicating autonomously. In this model, each “agent” could produce its “state,” and if that state changes, other agents subscribed to listen to it could make appropriate modifications (or prepare such changes for human approval before application). This is what I define as a “dynamic” approach, where all agents are interconnected and react to each other’s evolving states.
Naturally, such a dynamic approach presents numerous limitations. A primary concern is the significant increase in agent calls due to the interconnected and reactive nature of the system. However, the cost of LLM API usage has decreased so sharply that this might be manageable, at least for the foreseeable future. Another critical point is the challenge of making such dynamic agent communication truly efficient and reliable at scale. A2A, for instance, operates with “Agent Cards,” which reduce a subjective agent’s capabilities to an objective, discoverable text description, serving as a step towards standardisation in this dynamic environment.
Despite all this innovation, a conceptual gap persists in the community. We have agents, MCP, and A2A. On paper, everything seems ready for true intelligence. But the way most systems are built today, agents are often reduced to task-specific tools orchestrated by a central LLM “brain.” We call them agents, but they act like plugins. Their autonomy is usually constrained to rigid schemas and predefined tasks. This isn’t the intelligence many envisioned. It’s rather orchestration, not emergence.
Rediscovering the sculptor’s vision
As we stand at the threshold of the next era in computing, perhaps the breakthrough won’t come from better automation, but from rekindling the creative spark that inspired this field to begin with. It’s time to ask:
Do we truly need to confine complex human tasks, such as planning travel and purchasing tickets, to the sequential, often cumbersome, web-based processes of today, which demand constant manual input and navigation? Or can we fundamentally reimagine and execute these activities with far greater intelligence and efficiency, moving towards proactive, context-aware systems that leverage the full, interconnected capabilities of modern AI to anticipate needs and act autonomously?
This might require us to look beyond the immediate practicalities and embrace a more fundamental, perhaps even “old school,” approach to intelligence itself. One such idea involves reinventing multi-layer perceptrons (MLPs) with back-propagation, by replacing neurons with agents and tools. But that, dear reader, is a discussion for our next articles. So, stay tuned.
