Automated Workflow Architecture

API
1

Reading Serge Liatko’s fantastic post “Prompt Engineering Is Dead, and Context Engineering Is Already Obsolete: Why the Future Is Automated Workflow Architecture with LLMs” sparked some ideas about how to approach my AI development challenges. So I began working on a new model architecture, drawing on the concepts from that article, which I’d like to share with you here.

First of all, I must say that I’m not using any agent-based framework for development; I’m using plain Python. I have tried LangChain, OpenAI assistants, Botpress, AutoGen, and still have to check ADK, but I prefer the flexibility and freedom of Python and the OpenAI API.

What I have done:

Automated Workflow Architecture
Automated Workflow Architecture1077×948 85.6 KB

Let me try to explain the components of this model:

  1. System prompt + user prompt

  2. Workflow planner → generates workflow

  3. Worlkflow is composed of steps, some are fixed steps, some are AI Generated.

  4. The “Workflow execute” component loops through the steps passing Context*

  5. “Step execute”, executes each step

  6. If the previous fails, a mechanism regenerates and retries

  7. Finally the “Workflow execute” returns the output.

(*) Context is a JSON structure passed as a parameter to all steps and returned from any step.

There is still a long way to achieve the full architecture build, but I’m already getting some benefits from this approach:

  • Atomic steps are easier for the model to resolve → simpler questions return more accurate responses

  • Easy to find and fix issues, small issues → easier and faster fix

  • Automated context generation flows through all steps, feeding the steps with the info they need

  • Overall accuracy improved, lower risk of hallucinations

I still have a lot of testing to do. My plan is to recover logging errors in order to design a strategy for fixing recurrent issues. I’d also like to find a way to control output deviations (I’m considering a dual content generator agent supervised by a third agent). Finally, I want to implement a corrector agent that automatically detects error patterns, generates candidate fixes, learns from outcomes, and self-improves without manual intervention.

That’s where I’m at so far. I’d be glad to hear your thoughts, ideas, or even criticisms—always happy to learn from others tackling similar challenges

4 Likes
Prompt Engineering Is Dead, and Context Engineering Is Already Obsolete: Why the Future Is Automated Workflow Architecture with LLMs