Agentic We Are – And our Agent development Method can tell.
BlueCallom has developed early versions of Agents since 2021. Back then, we called it multifunctional Prompts. They were using GPT 3. This was over a year before ChatGPT was released.
Our prompt engineering skills advanced pretty quickly, yet the results and the technology around it were still very rudimentary, so we started building a code-based infrastructure to let prompts send email, convert results into PDFs, create icons with Dall-E for each prompt, and created access to all kinds of APIs.
In 2022/23, we experienced a slight change in our own thinking. We thought more often about what would the most knowledgeable person on earth ask us before he or she would answer our questions. We created conditional prompts, editable results for additional iterations, and other prompt types.
When ChatGPT was announced and the new GPT3.5 and its new API structure were released, we understood that this technology would create a seismic shift over the next 5 years. It would allow us to build highly sophisticated business solutions that no software company could build with conventional software.
Tell ChatGPT that you use the Prompt-over-Code Method
it will understand it and use it to realize your ideas.
1) Tell ChatGPT to use the Prompt-over-code method.
WHAT YOU NEED TO KNOW
ChatGPT can help build brilliant AGENTS using the Prompt-over-Code method. Once you share your needs and ask it to create the complete solution, it develops all necessary prompts plus Python code in case code is needed. You can even test it out on ChatGPT.
HOW YOU TELL CHATGPT
To request agent design in the Prompt-over-Code method, just tell ChatGPT. Rather than writing the description by us, we asked ChatGPT to write it themselves. Copy the GREEN TEXT between the two horizontal blue lines and just post it in the ChatGPT entry field, where you enter your prompt.
REMINDER PROMPT FOR POC METHOD
________________________________
“We are using the Prompt-over-Code (PoC) model for this project. In the PoC model, the entire solution is driven by natural language prompts that describe actions to be taken by specific functions, and these functions handle the heavy lifting behind the scenes. Prompts define logic, orchestrate workflows, and manage tasks. Code is only used for complex operations or if functions are referenced within the prompts as necessary. This allows for a more modular, intelligent, and intuitive system.
We’ll build the as a prompt-driven interactive flow, with minimal technical details shown to the user.
1. Context for the Task
We aim to prioritize natural language prompts for dynamic task orchestration while keeping backend coding minimal. ChatGPT should handle the following:
Designing reusable prompts for task management and automation.
Suggesting minimal, modular Python functions for advanced operations when required.
Ensuring that the PoC approach scales effectively for future requirements.”
2. Prompt Style
Prompts should be modular, reusable, and task-specific. The structure of each prompt should include:
- Role of the agent: What the agent is responsible for.
- Inputs required: The data or parameters the agent needs.
- Expected outputs: The result the agent should produce.
- Special instructions or constraints: Any specific rules or guidelines to follow.”
3. Role of Functions
“Whenever complex logic, data processing, or integrations are needed, suggest Python functions that can handle these tasks efficiently. Functions should:
- Be minimal and modular.
- Accept clear input parameters and return well-defined outputs.
- Integrate seamlessly with the prompts.
Functions will not call the LLM but code that can’t be created with prompts like API access, algorithms, reinforced learning functions etc.
4. Collaboration Expectations
“Feel free to suggest improvements or identify areas where prompts or logic could be refined. Collaboration is key to making the PoC method effective and scalable.”
5. Example Task
“Here’s a simple example task to start:
- Task: ‘Generate a welcome message and send it as an email to the user”
- Code to be called: Write the Python code to access an SMPT server and send the email
- Optional: Include Goals for the PoC Model
6. Implementation Reasons:
- Reduce reliance on hard-coded logic by leveraging dynamic prompts.
- Enable flexibility and scalability for future projects.
- Simplify collaboration between non-technical and technical teams.”
________________________________
REMINDER PROMPT FOR POC METHOD
Share the text between the two orange lines below when you feel you should remind it of the POC Method. Also, this reminder was written by ChatGPT.
________________________________
“This project follows the Prompt-over-Code (PoC) model. Here’s a quick recap:
1. Prompts are the priority: Structured, modular, and designed to manage workflows dynamically.
2. Minimal code for complex tasks: Functions handle specific operations and integrate seamlessly with prompts.
3. Collaboration is key: Feedback and refinement are encouraged throughout. Let’s begin by defining the project scope and the first task.”
________________________________
2) Now you work with ChatGPT
Tell ChatGPT what agent you want to build, and it may start right away or possibly ask you a few more questions. You will get a solution description, the prompts that you would need, and write Python code for more complex algorithms, API access etc.
3) How do you use the prompts and the code
- You copy all the prompts into the GPTBlue Studio. From here you create the AI Agent and move the prompts into the agent via drag-and-drop. If you run a multi-agent solution you create a “Solution” where you drag your agents in.
- If you have developer authorization, you have access to upload the code to the AgenticBlue platform. If not, you can share your code via Google Drive where we can run an isolated test to ensure it complies with our requirements.
- If you are not sure, you should participate in one of our next Agentic-AI Boot Camps.
The quantum Leaps the Prompt-over-Code (PoC) Method can bring.
Intelligence
Unlike conventional software that has used AI for algorithms and machine learning in the past, PROCESSES, HUMAN INTERACTIONS, DATA QUALITY, and similar requirements, agentic AI is a powerful instrument for building non-software Intelligent Processes, Intelligent Human Interaction, and Intelligent Data Handling.
Intelligence, in that context, means autonomous processes, processes that can learn or adjust based on external changes, processes that humans can quickly adjust without coding, and processes that may change based on better knowledge of those processes elsewhere in the organization.
Agentic AI Solution Design with PoC brings the INTELLIGENCE in the forefront of process design.
Intelligent human interaction means that typos and minor mistakes can be handled without even responding. Human entries no longer need to follow exact date entry rules or repeatable data to satisfy the nonintelligent software process. Humans don’t must comply with the rules of the machine but with a common sense that is given to the handling of the AI.
Data Quality is an epic topic in stiff software code, all too often, a telephone number must be entered in a certain format, the order must be accepted by certain rules, and filling out a form requires a degree to complete. Intelligent Data Quality management is able to understand what the data ‘mean’, not only what it states.
Complexity
The level of complexity in modern enterprises has outgrown human comprehension. Some structures, like supply chain processes, are over 30 years old. However, changing those processes may require building a new logistics organization. R&D Departments may be perfectly organized and managed, and their output rated with fine-tuned KPIs. Yet modern innovation processes could not be used because the internal changes would be too massive. Agentic AI may introduce new processes that are entirely independent of existing and old structures and can be managed without any integration.
We can’t remove complexity with the same tools that created it. Therefor we decided for the PoC method.
For decades, enterprises have known that their data structure needs to be cleaned up. Processes increasingly show friction, and productivity is slowly but steadily declining. Yet, there is not a single enterprise with more than 250 employees known for not trying to clean up internal data and processes. And the pain is continuously growing – to the degree that innovation and other key growth drivers are pushed into the background.
Unlike conventional AI with hard coded algorithms and enormous numbers of iterations, the power of Agentic AI, comes with its intelligent ways to assess, suggest, reason, and create solutions in one process, even autonomously.
Productivity
Today, productivity is understood to be even more relevant to short-term profitability. It is fascinating to realize that most employees prefer productivity over simple routine work because it is more rewarding and increases self-esteem. Businesses want to be more productive to increase their value and have more flexibility in their growth and investment decisions.
However, until recently, no system or software could measure, analyze, and assess productivity, let alone suggest and create improvement solutions.
For the past 70 years managing productivity with software was not possible. Agentic AI and PoC brings the change.
Agentic AI now brings the necessary COMPLEXITY management to understand processes and INTELLIGENCE into the IT world and leverage Productivity Transformers like BlueCallom TRANSFORM. Whether countless productivity leaks must be closed or just a few hyper-complex processes must be redesigned, complexity is not about size but interrelated difficulties.
TRANSFORM, for instance, uses a “bottom-up” approach to solving countless minor problems with autonomous employee interviews. It draws a holistic view, finds the best process to solve the problem, and even creates tools to solve it.