A quick message from us at 1752vc (Formerly Pegasus):

Accelerate: Where Traction Meets Velocity

Our highly personalized three-month program offers $100K in funding, expert mentorship, and a powerful network to help you achieve your business goals.

Apply Here

In the first article of this series, we introduced the idea that large language models represent only the first step in a much broader shift toward agentic AI. Models that generate text, code, and images have captured public attention, but they are only one component of a larger architecture that will define the next phase of artificial intelligence.

At the center of that architecture sits cognition.

Cognition, in this context, refers to the set of processes that allow an intelligent system to learn from experience, interpret information about the world, reason through possible outcomes, and decide what to do next. Humans rely on cognition constantly. Every decision we make is shaped by what we remember, how we interpret the present moment, and what we believe might happen in the future.

AI systems are beginning to develop analogous capabilities.

Modern agents are no longer designed merely to respond to a single prompt in isolation. Instead, they operate within systems that maintain internal state, track goals, retrieve information, and generate sequences of decisions over time. The moment AI systems begin interacting continuously with environments—digital or physical—they require the ability to learn, reason, and adapt.

Understanding cognition therefore means understanding the core processes that allow agents to move beyond static outputs and toward dynamic behavior.

Two processes sit at the center of this capability: learning and reasoning.

Learning allows a system to update its internal representation of the world based on experience. Reasoning allows the system to use that knowledge to determine what actions should be taken.

Together, these processes form the cognitive loop that underlies intelligent behavior.

Cognition as an Internal State

One way to understand cognition is to think of it as the maintenance and evolution of an internal state.

At any moment, a human carries a mental representation of the world. This representation includes what we believe to be true, what we remember from past experience, what goals we are pursuing, and what we expect might happen next. When new information arrives, it is interpreted within that existing context.

The mind is therefore not merely reacting to external inputs. It is constantly updating an internal model of reality.

Several subsystems contribute to maintaining this state. Perception gathers signals from the environment. Memory stores and retrieves past experiences. Reasoning processes evaluate possibilities and make predictions about outcomes. Motivation and emotion shape which possibilities receive attention.

These components work together to maintain a coherent view of the world that guides action.

Agentic AI systems are beginning to operate in a similar way. Instead of neurons and biological processes, they rely on neural networks, memory stores, reasoning mechanisms, and environmental interfaces. But the underlying principle is the same: the system maintains an evolving representation of the world and updates that representation as it gathers new information.

Every time an agent learns something or reasons through a decision, it is operating on that internal state.

Learning: How Agents Improve Over Time

Learning is the process through which an intelligent system improves its performance through experience.

Every interaction an agent has with its environment produces information. Sometimes that information confirms the agent’s expectations. Sometimes it reveals errors in the agent’s internal understanding. In either case, the experience can be used to refine the system’s internal model.

Over time, this process allows the system to make better predictions and better decisions.

Different learning mechanisms exist, but they all share a common objective: reducing the gap between the system’s internal representation of the world and the way the world actually behaves.

Some learning methods rely on labeled examples, where the system is shown the correct answer and adjusts its internal parameters accordingly. Others rely on pattern discovery, where the system identifies structure within large volumes of data without explicit supervision. Another class of learning occurs through trial and error, where actions that lead to successful outcomes are reinforced and unsuccessful ones are discouraged.

Despite their differences, these approaches can be understood as variations of the same underlying process: updating internal beliefs based on experience.

Learning therefore changes the cognitive state of the system.

The Balancing Act of Learning

Learning is not simply about absorbing information. It involves managing several competing forces.

One of these forces is adaptation. When the environment behaves differently than expected, the system must revise its internal understanding. Without this ability, learning would never occur.

Another force is exploration. If an agent always repeats behaviors that have worked in the past, it may never discover better strategies. Intelligent systems must occasionally try new actions in order to discover information that cannot be obtained through repetition alone.

A third force is stability. Learning too aggressively can cause systems to forget useful knowledge acquired earlier. This phenomenon, sometimes described as catastrophic forgetting, reflects the difficulty of updating knowledge without destroying previously learned information.

Balancing these forces—adaptation, exploration, and stability—is one of the central challenges in building intelligent systems.

Learning Across the Agent Architecture

In modern AI agents, learning does not occur in a single location.

Sometimes learning modifies the core model itself, improving the system’s ability to interpret language, recognize patterns, or perform reasoning tasks. These updates affect the entire cognitive system.

In other cases, learning occurs within surrounding components.

Agents may store experiences in external memory systems, allowing past interactions to influence future decisions. They may update reward models that guide decision-making. They may refine internal representations of how actions influence outcomes in their environment.

In many cases, adaptation happens through context rather than permanent parameter updates. By retrieving information, incorporating feedback, or storing new experiences, an agent can adjust its behavior without retraining the underlying model.

This flexibility allows agents to adapt to new environments far more quickly than static models.

Learning to Perceive Better

Perception is often treated as a separate problem from cognition, but in practice learning and perception are deeply intertwined.

An intelligent system’s reasoning capabilities depend on the quality of the information it receives about the world. If perception is limited or inaccurate, even sophisticated reasoning processes will produce poor results.

For this reason, advances in perception often translate directly into improvements in cognition.

One of the most important developments in modern AI has been the expansion of systems beyond a single form of input. Early models operated largely within text, constructing internal representations of the world based on linguistic patterns. While powerful, this approach limited the system’s ability to interact with environments that communicate information in many forms.

More recent systems are increasingly capable of integrating multiple modalities. Images, audio, documents, and structured data can be interpreted alongside language. As agents incorporate these additional streams of information, their internal representations of the world become more grounded and more detailed.

Perception is also evolving through the integration of information retrieval.

Traditional models rely heavily on knowledge encoded during training. That knowledge can be extensive, but it remains static once training ends. Retrieval systems alter this constraint by allowing agents to access external information sources during operation.

Instead of relying exclusively on what they remember, agents can search documents, query databases, or gather information from external environments in real time.

This transforms perception into an active process.

The agent is no longer limited to the knowledge contained within its internal parameters or the immediate information provided by a user. It can dynamically expand its understanding by acquiring new data relevant to the task at hand.

Over time, this capability may evolve further. Rather than simply retrieving information when prompted, agents may develop the ability to recognize when their knowledge is insufficient and proactively gather additional information.

In that sense, perception becomes not just the interpretation of inputs but the navigation of an information landscape.

Learning to Reason Better

If perception determines what enters the system, reasoning determines how that information is transformed into decisions.

Reasoning is therefore one of the most critical capabilities in agent cognition.

In recent years, research has increasingly focused on improving reasoning performance directly. A wide range of strategies have emerged, many of which attempt to teach systems how to think through problems rather than simply predict answers.

One approach emphasizes the importance of structured reasoning examples. Instead of learning only the final outputs associated with a problem, systems are exposed to step-by-step explanations of how solutions unfold. Observing these reasoning traces helps models develop internal representations that support more reliable problem solving.

Another approach focuses on iterative improvement. In these systems, the agent generates potential reasoning paths, evaluates their quality, and then learns from the most successful ones. By repeatedly generating, evaluating, and refining reasoning strategies, the system gradually improves its ability to solve complex problems.

Evaluation mechanisms also play a role. Rather than committing immediately to a single answer, some systems explore multiple candidate reasoning paths and compare them before selecting a final result. This search-like process often improves performance on problems that involve multiple steps or uncertain outcomes.

Reinforcement-based methods extend this concept further. In these frameworks, reasoning trajectories themselves are evaluated according to whether they lead to successful outcomes. Feedback signals then guide the system toward more effective reasoning strategies.

These developments suggest that reasoning is not a fixed capability embedded within a model but a skill that can improve through experience and feedback.

Structured Reasoning: Thinking in Steps

One of the most influential discoveries in modern AI reasoning research is that models often perform better when encouraged to reason explicitly through intermediate steps.

Rather than producing an immediate answer, the system generates a sequence of logical stages that lead to the final conclusion.

This structured approach helps models manage complex tasks by breaking them into smaller pieces. It also provides a level of transparency that allows observers to inspect how the system arrived at its answer.

Structured reasoning methods have evolved significantly from this basic idea.

Some techniques focus on automatically generating useful reasoning examples. Others explicitly decompose complex tasks into simpler subproblems that can be solved sequentially. Still others encourage the system to first consider high-level abstractions before moving into detailed reasoning.

Over time, these techniques have expanded into more elaborate reasoning frameworks. Instead of following a single chain of thought, models can now explore multiple reasoning paths simultaneously. Potential solutions are generated, evaluated, and refined before a final decision is made.

In effect, reasoning becomes a form of search.

Static and Dynamic Reasoning Structures

Not all structured reasoning operates in the same way.

Some approaches rely on static structures. In these systems, the reasoning framework remains fixed while the content varies. The model might generate several independent reasoning chains and then combine their conclusions. Because each chain approaches the problem differently, errors in any single path are less likely to dominate the final result.

Other approaches introduce iterative reflection. The system produces an answer, critiques it, and then attempts to improve it through additional reasoning cycles.

Dynamic reasoning structures go further. Instead of following a predetermined sequence of steps, the reasoning process itself can evolve as new information becomes available.

In these settings, the agent may alternate between thinking, acting, observing outcomes, and revising its strategy. Each action produces new information that reshapes the reasoning process.

This dynamic structure more closely resembles real-world cognition, where decisions rarely follow perfectly linear paths.

Unstructured Reasoning: Thinking Beneath the Surface

While structured reasoning has produced significant gains, not all reasoning requires explicit scaffolding.

Some approaches allow the system to rely more heavily on internal representations. Instead of spelling out every intermediate step, the model reasons implicitly and produces an answer that reflects a more holistic internal process.

Prompting techniques can still guide reasoning, but the structure is less rigid. The system is encouraged to consider problems more flexibly rather than following a predefined reasoning template.

Recent work has also explored reasoning processes that occur partly within latent representations rather than fully expressed language. In these settings, the system may generate internal rationales that influence its predictions without exposing every intermediate step externally.

These approaches remain an active area of research, particularly because implicit reasoning introduces challenges for interpretability and verification.

Nevertheless, they point toward an important insight: intelligent reasoning does not always resemble a written explanation.

Some forms of cognition occur beneath the surface.

Planning: Cognition Extended Through Time

Reasoning becomes planning when decisions extend beyond single steps and unfold across sequences of actions.

Planning involves moving from an initial state toward a desired future state through coordinated actions. This ability is a hallmark of advanced cognition because it requires the system to imagine possible futures before they occur.

Humans rely on planning constantly. We evaluate long-term consequences, consider alternative strategies, and adjust our behavior in response to new information.

AI agents are beginning to develop similar capabilities. They can decompose complex goals into smaller tasks, generate possible strategies, and evaluate alternative pathways toward an objective.

However, long-horizon planning remains difficult for many systems. Models that excel at short reasoning tasks often struggle to maintain coherence across extended sequences of decisions.

Research in agent planning therefore focuses on several strategies.

One approach involves task decomposition, where complex objectives are broken into manageable subgoals. Another relies on search processes that explore multiple possible action sequences. A third integrates environmental knowledge so that plans are grounded in realistic expectations about how the world behaves.

Planning thus sits at the intersection of learning, reasoning, and world understanding.

The Biggest Open Problem in Agent Cognition

Despite rapid progress in reasoning research, one critical challenge remains.

Much of the field’s attention has focused on improving the mechanics of reasoning—how systems construct logical chains or explore alternative solutions. But reasoning alone is insufficient.

An agent can reason perfectly within its internal logic and still fail if its understanding of the environment is inaccurate.

This highlights the importance of world models.

If a system lacks accurate knowledge about how actions influence outcomes in real environments, even elegant reasoning processes may produce poor decisions.

Increasingly, researchers are recognizing that improvements in reasoning must be paired with improvements in environmental understanding. Agents must be able to gather current information, simulate possible outcomes, and revise their internal models based on interaction with the world.

In other words, the future of AI cognition will likely depend not only on better reasoning structures but on deeper integration with perception, memory, and world modeling.

Better thinking alone is not enough.

Systems must also think about the world as it actually behaves.

What Comes Next in This Series

This article has focused on the cognitive core of intelligent agents. But cognition does not exist in isolation.

Once an agent can learn and reason, a broader set of architectural questions emerges. How should experiences be stored so they can influence future decisions? How should an agent model the environment it operates in? How should reward signals shape long-term behavior? What happens when agents interact with tools, software systems, or one another?

The rest of this series explores these questions.

The next articles expand outward from cognition into the full architecture of agentic systems—examining memory, world models, action systems, self-improvement mechanisms, and eventually networks of interacting agents.

Taken together, these components form the emerging stack of intelligent systems.

Large language models may have started the story, but cognition—learning, reasoning, and planning—is what begins to turn models into agents.

Series Note: Derived from Advances and Challenges in Foundation Agents

This series draws heavily from the paper Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems (Aug 2, 2025). The work brings together an impressive group of researchers from institutions including MetaGPT, Mila, Stanford, Microsoft Research, Google DeepMind, and many others to explore the evolving landscape of foundation agents and the challenges that lie ahead. We would like to sincerely thank the authors and researchers who contributed to this outstanding work for compiling such a comprehensive and insightful resource. Their research provides an important foundation for many of the ideas explored throughout this series.