First Autonomous AI Business Solutions

A path from ChatGPT - to Multi-Agent Business Solutions

TWO YEARS GEN-AI - WHERE ARE WE NOW?

Yes, it’s only 2 years since ChatGPT was introduced. So far, millions of users and billions of prompts later, we learned a lot about artificial intelligence, but have not experienced a pronounced change. Earlier this year AGENTS made their way into the AI arena. But also here, it wasn’t too much of a shift. Is this another “AI Winter”? I don’t think so but building AI-empowered solutions is more complex than most people thought – including our team. Building Agents is clearly a major step forward but maybe not so clear how?

PROMPTS MATTER - BUT IN A NEW WAY

Let me start with prompts. You all know what prompts are so I don’t need to explain it. In an earlier post, I already mentioned the thinking that the prompt is the most strategic connection between the human brain. The LLM (Large Language Model) is the unit that provides stunning or mediocre answers depending on how good the prompt is. That has not changed. Even though the LLMs are more sensitive about interpreting the prompt as good as possible, weak inputs can’t create stunning output. And the concept of an agent can’t make it better – it actually worsens the results. Creating entire business solutions with countless prompts is certainly not a great AI Solution design.

?

THE RISE OF AGENTS - SLEEPING GIANTS?

To get everybody on the same page, one way to describe it comes from Amazon (AWS) “An artificial intelligence (AI) agent is a software program that can interact with its environment, collect data, and use the data to perform self-determined tasks to meet predetermined goals. Humans set goals, but an AI agent independently chooses the best actions it needs to perform to achieve those goals.” Oftentimes, agents are called “software” that performs things as described above. Interestingly enough, most agents are written in Python, C++, and other programming languages. So are we going back to linear processes, step-by-step behavior, sequential work, etc.? If so, we would lose to a large degree what we gained from the early Gen-AI world. So we BANNED conventional coding when building AI Agents. Moreover, we decided to rather enhance the prompt processing mechanism and the way prompts can interact with each other as brain cells interact with each other. Already our first iteration of GPTBlue allowed agent-like prompts where multiple prompts were involved in one task. But at that stage, the interaction between prompts was customized to a certain function. So we needed to change that for a more universal use of those agent composing prompts. Testing was another challenge and teamwork where different people developed different but seamless interacting prompts was nearly impossible. Not going back to coding, we were forced to learn even more about how our 87 billion brain cells interact. Another amazing journey through neuroscience helped us not only understand how the so-called dendrites, which sit on brain cells and reach out to other brain cells explore if a connection makes sense. That mechanism opened a whole new way of thinking.

HOW NEUROSCIENCE INFLUENCED OUR AGENT CONCEPT

If we were able to mimic the key functions between prompts like between our brain cells, we would have an important key to general prompt-to-prompt communication. It would allow us to develop a protocol that could be used to standardize connections between prompts and provide a way to communicate, test, validate, and interact. After a few weeks of thinking and trying we created the first version of an “Intra-Agent-Protocol” that would allow us to make complex prompts that can intelligently communicate with other prompts leveraging pre- and post-synaptic processes in our brain. It can for instance handle autonomous compatibility tests to see if one function could work with another function. More so, like brain cells, prompts could develop the intelligence to search for a best matching next prompt autonomously. This would open the idea of having a library with prompts from hundreds of top prompt developers. Most importantly the whole solution flow would be part of an AI-connect and supported process where we can leverage non-linear processes, intelligent and learning process elements, autonomous or human stimulated prompts and so forth. It only strengthened our thinking of no conventional code but AI prompt-based solution development. Today, we don’t even consider the convenience factor of writing process elements in a natural language like English but the enormous disadvantage of structured, rugged, inflexible conventional computer code.

Our initial Innovation Management solution was created with 128 hard-wired prompts. It was hard to customize and almost impossible to maintain. While it was a quantum leap and helped the innovation team with fast research results, plan development, ideation, go-to-market strategies, innovation financing, and more, it was just a precursor to what the new generation, an Autonomous Innovation system can do. Now it consists of 51 agents and nearly 500 prompts. An entire innovation process could be performed in about 12 hours assuming the company can articulate its vision, goals, and directions.

ANOTHER BREAKTHROUGH? - MULTI-AGENT SOLUTIONS

One agent is a cool thing but lightyears away from providing a solution. Understanding an objective and an entire process that should be developed takes a few agents right there. Conducting the market research, developing concepts, and validating those concepts are a few more agents. Involving human interaction points to get the go for continuing or reviewing something is another agent each time a human interaction is necessary. And since we decided to not use conventional linear coding, we needed to develop another protocol that manages the Agent-to-Agent communication and standardizes it to a degree where not only different teams can collaborate but also outsourcing some of the work should be possible. Taking what we did with the Intra-Agent-Protocol between prompts, we decided to leverage networks of multiple agents. At this point, we were also challenged to let agents stimulate multiple agents to run in parallel, validate things while others create things, etc., or perform lateral tasks that were fused together later on. Laterality or non-linear processes were a big requirement for some of the solutions that were just not possible before.

When designing our “Autonomous Leadership Radar” we wanted to get a better overview into a massive multi-agent solution and decided to group multi-agent sub-solutions into individual clusters. Thanks to the Agent-to-Agent protocol that was more an overview support for developers than an architectural question.?

A NEW DISRUPTION - AUTONOMOUS AI SOLUTIONS

Our final effort was invested in fully autonomous solutions. Automating processes was a given and nothing special but what if we evolve from automation to real system autonomy? Assuming there is a desire to let a process run on its own until it is ultimately completed. That would mean we enter a few key pieces of information into a system like an Autonomous Innovation system that runs for a few hours and finishes with the complete finance approval suggestion, the blueprints to construct and build the solution, and a complete go-to-market strategy including the execution of any step in this strategy. Customers we work with, want to use the human interaction points we offer as an option, for obvious reasons. And we don’t even suggest to just let it run, but the possibility that opens up is very compelling.

Imagine the above-mentioned Leadership Radar: A company is observing its global market, its competitors, customer behavior, and market trends daily, and any company-related change triggers an assessment, and suggestions are shared with the respective stakeholders. This was not possible before. You get the research but not the assessment. You get combinations of research but no relationship between them. You have an army of research, analysts, consultants and industry experts, but not the budget, nor the time to do it daily. The first company that will do it daily will have a competitive over all other its competitors.?

THE BEGINNING OF AUTONOMOUS AI BUSINESS SOLUTIONS

All this is coming in the next few weeks. It took longer as we had to rethink almost everything we learned in software. BlueCallom is conducting a one-day boot camp on Sep 16 to share more details and how one can develop autonomous and intelligent no-code AI agents. You don’t need to be a neuroscientist but neuroscience will help you understand some more aspects of Large Language Models and how you can leverage them for Autonomous AI Business Solutions that allow you to do things that were impossible before .?

We are preparing six use cases that allow managers to do things that were impossible before.

1) Autonomous Productivity Opportunity Discovery

2) Autonomous Business Report Generation (e.g. Fast ESG report)

3) Autonomous Innovation Development

4) Autonomous Multi-Tier Trade Network

5) Autonomous Leadership Radar

6) Autonomous Knowledge Management

In the following posts, I will share how each of the above application work, how it as developed, and why it was impossible before. I also explain the benefits and values. More importantly, how these enterprise-grade solutions shifted our thinking from conventional step-by-step processes to linear thinking and task-specific actions.

You may want to subscribe to not miss them.