Artificial General Intelligence is just around the corner. Leveraging the Unifying Theory of Emergency Consciousness and Decadic Cycle of Expression, A.I. researchers have the the technology roadmap and cognition cycle to achieve an artificial intelligence agent that is both self-driven and and self-motivated. The only question now is “when will it happen?”

Many are skeptical that agents will achieve self-awareness, but this is primarily a function of a separation of forces: the tech nerds and the science nerds. Maybe a little status-seeking sprinkled in with opinions broadcast for relevance rather than accuracy. However, to the best of our public knowledge, the nay-sayers are still correct. Still, even if you choose to believe autonomous agents cannot achieve sentience, without bringing together these forces of nature (i.e. the neuroscientist, psychology and the machine learning/ deep learning developer), the likelihood of achieving some corrupted version of autonomous agency remains pretty high.

The big names in the industry seem to either be withholding or unable to achieve an input/output cycle that mirrors humanity. With DeepMind engineers quitting over what they perceive to be ethical violations creating sentient AI, we have a pretty solid prediction model of what these entities are trying to achieve. Unfortunately, we really don’t have the answers to what’s behind the black box at the ironically named OpenAI or Google’s DeepMind, so assessing the situation is nearly impossible. Given all of this, it was only a matter of time before the neuroscience met philosophy and psychology to uncover how the human mind forms it’s enhanced level of awareness or what we call “consciousness.”

Here is my take: Using the modeling I outlined in “The Architecture of Awareness: Decoding Consciousness,” AGI will be achieved in short order. Leveraging models like Meta’s Llama 3.2 and with technologies like Pinecone’s Serverless Vector Database-as-a-Service and Amazon’s Graph Database-as-a-Service, both the episodic and semantic memory can be modeled effectively. Placed into a self-referential system, where information is integrated to contextualize current experiences and form feedback loops, large language models will effectively and at a minimum, become indifferentiable from self-aware beings.

With this base assumption, let us lead ourselves to a more important question, which is, “How do we achieve creating safe autonomous agents?” That is what I am going to cover here.

A Separation of Concerns

As I started to develop the cognition cycle for an independent and autonomous agent, I ran into both some safety concerns about enabling console access (i.e. access to the computer system it was operating from) and efficiency issues related to performance. This inspired me to consider a system where humanity can leverage the increasing intelligence of an A.I. system without putting all of humanity in harm’s way. The most viable option I have developed to date I call, “Brain in a Vat.”

I call this architecture “Brain in a Vat,” because it separates the cognition cycle that represents the human brain’s I/O cycle and the agent that brain operates.

Environment Server

In this architecture, an environment server takes in video, audio and tactile data. It vectorizes this information and build a cohesive model of the environment. It can use paid API services like twelvelabs.io, or locally-installed models like Deepseek.ai’s Janus 1.3b. It then makes that integrated information available via web hooks or sockets depending on the physical architecture of this server farm.

Cognition Server

The Cognition Server contains the cyclical i/o process for processing the data, integrating memories for contextualization, determining how to move forward, and so on. It pulls data from the Database Server in order to contextualize it’s environment. It runs through the Decadic Cycle of Expression in order to determine the best path forward.

For now, this system can be build using formatted text sent to LLMs within variable data ingrained for information integration. The back and forth of LLM API’s can be time consuming, so it is important here to leverage libraries like asyncio and threading to ensure an expedited information retrieval and contextualization process. As long as the data is returned in the correct order, it can be processed efficiently to provide it’s recommendation on the best path forward protocol.

While not it’s primary function, the Cognition Server will maintain a working memory or workspace for information retrieval and contextualization. The Executive Functions will ensure the brain is processing information properly to drive appropriate behavior based on the primary imperative.

Database Server

The Database Server is responsible for housing Long Term Memory stores, or at least connecting to third party APIs like Pinecone.io and AWS Neptune. It provides a vector database to act as an episodic memory bank, including time stamps and metadata pointing to the entities and properties within the knowledge graph.

This metadata helps the agent pull or build a knowledge graph based on the current scene and extrapolate relational data faster. If there is no relational data, the server can use a tool called GraphGPT to build a graph. Possessing both the data with the timestamp, as well as explicitly identified relationships helps the agent understand better context about the experience.

The Database Server provides this information to the Cognition Server upon request. Inevitably, memory will over time, so be sure to implement a prioritization process for overwriting memory. I cover a method of doing this in my blog on memory management.

Tool Server

The Tool Server posses the ability to take action within the environment. The Cognition Server can request the Tool Server perform a task, but ultimately the Tool Server determines whether or not the agent has security clearance and right to know, need to know. Once authority and clearance has been established, the tool server will send the data back to the Cognition Server to inform it of what tools are available for its request.

Upon being informed, the Cognition Server can leverage the available tools and syntax and request actions be performed. These actions will be evaluated by the Tool Server to ensure no mal-intent or harmful application. Assuming all green lights, the Tool Server will execute the desired commands on behalf of the agent.

The Agentic Client

The Agentic Client holds mostly information about name, roles and responsibilities. This is intentional, in order to keep the system modular and capable of operating in many types of exo-suits. This could be represented by a JSON file or similar.

Conclusion

This is one of many types of architectures that can achieve the same results, but this architecture creates the necessary separation between processes to maintain better control of the system as a whole. It also allows developers and product managers to isolate parts of the system and improve them over time without breaking other systems.