top of page
Search

Coding the Mind: How Building an AI Agent Mirrors the Architecture of Human Thought

  • Writer: Adri Research Forum
    Adri Research Forum
  • Apr 17
  • 7 min read

By Bedabrata Bagchi | April 2026 | ARF




The Premise:

There is something unmistakably familiar about the way a modern AI agent works. It receives a stimulus. It weighs it against memory. It forms a plan. It acts. It observes the result and corrects itself. If this sequence sounds biological, that is because it is. The architecture powering today's most capable AI agents is not merely inspired by the human brain in its structural logic, it is a deliberate translation of how the human mind processes thought, step by measurable step.

This article makes that case, grounding it in recent research and the hard numbers behind both biological cognition and computational intelligence.

**(Disclaimer - These parallels are architectural analogies, not claims of equivalence.)

Part One: The Brain as the Original Agent

The human brain is arguably the most sophisticated information-processing system ever studied. It comprises over 86 billion neurons, each capable of forming up to 10,000 synapses with other neurons, resulting in an exceptionally complex network of connections (Source). At the macro level, the brain contains over 100 trillion synapses, storing and processing information much like a computer, where the brain matter acts as the hardware and the transmitted information as the software. (Source: IMR Press)

But what makes the brain particularly relevant to understanding AI agents is not its raw scale but more so its architecture of specialisation. Key brain regions are mapped to distinct cognitive functions: the prefrontal cortex (PFC) oversees advanced reasoning, planning, and executive function; the hippocampus manages memory and complex data processing; the basal ganglia and thalamus handle decision-making and adaptive learning; and sensory cortices manage basic pattern recognition and input processing. (Source: IMR Press)

Executive function, housed primarily in the prefrontal cortex, acts somewhat as the brain's air traffic control. Managing which tasks get resources, when attention shifts, and how conflicts are prioritised. (Source: Frca) This precise specialisation is the brain's version of what software engineers call "modular design."

The brain also operates with remarkable efficiency. It consumes only about 20 watts of power which essentially is equivalent to a dim light bulb. Which basically is accounting for 20% of the body's total energy expenditure despite comprising only 2% of body mass. (Source: Birow)

Part 2: Thought as a Function - The Step-by-Step Mechanism

One of the most profound realisations in cognitive neuroscience is that human thought is not abstract. it is a sequence. The brain does not arrive at a decision all at once. It processes input, consults memory, maps possible actions, selects one, executes it, and then evaluates the outcome. This is the cognitive loop.

Short-term memory, sustained through persistent neural activity in the prefrontal cortex, allows the brain to temporarily store information for seconds to minutes for immediate processing - typically around 7 ± 2 items, per Miller's Law. Working memory extends this by actively manipulating information for reasoning, decision-making, and problem-solving, integrating inputs from the PFC, parietal cortex, and hippocampus. (Source: arXiv)

Researchers using PET and fMRI have found that the prefrontal cortex is a key part of working memory, and this discovery directly inspired AI researchers to incorporate memory modules into machine learning models - one representative method being LSTM, which laid the foundation for many sequential processing tasks such as natural language processing and time-series analysis. (Source: Engineering)

What this tells us is that the very act of "thinking through a problem" - which we take for granted as humans - is a structured, hierarchical pipeline. And when developers sit down to build an AI agent, that is precisely what they are formalising: a pipeline of cognitive-like functions.

Part 3: The AI Agent as a Computational Thought Process

An AI agent is, at its core, a formalisation of cognition. The architecture of an autonomous agent is described as a cognitive pipeline that turns perception into action through a central decision-making process and integrating modular resources with a reasoning engine that runs over time.

It breaks into three layers:

  • Core Components (perception, memory, action, and profiling);
  • the Cognitive Architecture (planning and reflection); and
  • Learning.

(Source: arXiv)

The parallel to the brain's own structure is not coincidental. It is architectural.

AI development draws inspiration from the architecture and functions of the human brain, mapping neural and computational architectures to show how AI models progressively mimic the brain's complexity, from basic pattern recognition and association to advanced reasoning. (Source: IMR Press)

The dominant reasoning paradigm in modern agents - the ReAct framework (Reasoning + Acting) - is particularly instructive here. ReAct enables language models to generate both verbal reasoning traces and text actions in an interleaved manner. Reasoning traces affect the internal state of the model by reasoning over the context and updating it with useful information to support future reasoning and acting. (Source: Google Research)

In practice, the loop runs: Thought → Action → Observation → Thought again. This dynamic looping closely mimics how a human problem-solver alternates between thinking and doing, adjusting plans on the fly. (Source: Wollenlabs)

The performance lift is measurable. ReAct with one-shot and two-shot prompting outperforms imitation and reinforcement learning methods trained with approximately 100,000 task instances, with an absolute improvement of 34% in success rates on ALFWorld and 10% on WebShop over existing baselines. (Source:Google Research)

Part 4: Function Definitions as Externalised Cognition

Here is the core insight of this article: when a developer writes an AI agent, they are not writing code in the traditional sense. They are externalising and defining the implicit functions of thought.

Consider what happens when a human expert approaches a problem. They perceive the context. They retrieve relevant knowledge from memory. They identify what "tools" they have available - data, contacts, expertise. They form a hypothesis. They test it. They update their model. Every one of these micro-actions, in the human brain, is carried out by a specific neural system. In an AI agent, each of these becomes a literal function.

The ReAct framework exemplifies this by combining reasoning (chain-of-thought prompting) and action (tool use), with LLMs alternating between internal cognition and external environment interaction. Reasoning in agents like AutoGPT is enhanced with function-calling and prompt chaining to simulate thought processes - for example, "step-by-step" prompts or intermediate tool invocations. (Source:ScienceDirect)

A defining feature of agentic AI is the latent reasoning step - "the thought or plan" - that is produced before any external action is taken. This variable can represent a simple chain of thought or a more structured hierarchical plan. (Source: arXiv)

This mirrors what neuroscience has long observed in the PFC. The prefrontal cortex is responsible for decision-making, regulating planning, execution, and task management. It can be divided into multiple core modules, including the Dorsolateral PFC for working memory, the Ventromedial PFC for risk assessment, and the Anterior Cingulate Cortex for error monitoring and emotional regulation. Each module in the brain corresponds to what a developer might call a distinct sub-agent, a tool, or a prompt chain.

Part 5: Brain-Inspired Architecture - The Research Frontier

The scientific community is actively closing the loop between neuroscience and agent design. Recent studies have shown a strong similarity between Artificial Neural Networks (ANNs) and Biological Neural Networks (BNNs), prompting many researchers to devote to this field. Agent architectures have incorporated elements that mimic brain structures - such as the Talker-Reasoner dual-system architecture, which simulates the brain's fast and slow thinking systems. (Source: arXiv)

At Johns Hopkins University, researchers have found that architecture - not data volume - may be the key differentiator. Untrained convolutional neural networks, when modified to better simulate patterns in the human brain, rivaled conventional AI systems generally exposed to millions or billions of images during training, suggesting that architecture plays a more important role than researchers previously realised. Hub As the lead researcher observed: "Evolution may have converged on this design for a good reason. Our work suggests that architectural designs that are more brain-like put the AI systems in a very advantageous starting point." (Source: Hub)

Meanwhile, in the domain of LLMs and language specifically, large language models have emerged as a new candidate model for the study of language, exhibiting human-level performance on diverse language tasks, with internal representations that are similar to the representations in the human brain when processing the same linguistic inputs. (Source: PubMed Central)


Section 6: Numbers That Tell the Story


Agents operating in multi-agent systems consume approximately 4-15x more tokens than standard chat interactions, reflecting the cost of sustained cognitive loops. (Source: Oracle)


Conclusion

Building an AI agent is, at its most honest description, an act of cognitive cartography. Developers are mapping the terrain of human thought - perception, memory, planning, action, reflection - and encoding each landmark as a function. The brain does this in electrochemical signals across 100 trillion synaptic connections. The agent does it in prompt chains, tool calls, and memory stores.

The gap is enormous, and the humility is warranted - the brain remains roughly 8,000 times more structurally complex than today's largest neural networks. But the architecture of cognition, which essentially is the logic of how thought moves from input to reasoning to action to correction, is something we have now learned to replicate in code.

The implication is significant for anyone building agents today: the closer your design mirrors the modular, iterative, reflective structure of human cognition, the more capable your agent becomes. That is not a metaphor. It is, increasingly, the empirical finding of both neuroscience and AI research.


References
 
 
 

Comments

©2026 | by Adri Consulting Group. | Reg: Adree Consultants | India

bottom of page