Scalable Agentic AI Systems and the D3 Framework

The expertise driving these insights comes from an individual who served as the Head of AI at Amazon Germany, responsible for building large-scale AI systems that generated billions of dollars in revenue impact.
Professional experience extends into robotics AI, leading to the creation of a startup called Omokai.
Omokai is developing a physical AI operating system designed to convert voice commands into autonomous missions for robots and drones.
The organizational structure at Omokai is AI-native, employing 7 human staff members and more than 10 AI employees.
The technological stack at Omokai utilizes custom, fine-tuned models and leverages AI to generate data.

Research published by Gartner indicates that increasing the use of AI do not inherently lead to increased productivity; instead, unbounded AI usage without proper constraints often results in chaos and noise.
There is a significant amount of hype surrounding "agentic systems," leading to several detrimental industry myths:
- Agent Swarms: Building systems with hundreds of agents to solve a single problem often leads to exorbitant costs and a high frequency of errors.
- Loop Engineering: Implementing "loops within loops" significantly increases operational costs without guaranteed improvements in outcome.
A common misconception is that AI will become so inexpensive in the future that these costs will not matter.
Uncontrolled AI usage in code generation results in persistent errors within production databases, creating long-term technical instability.

Beyond direct financial expenses, using AI without control mechanisms introduces several hidden costs:
- Verification: The time and resource requirements to verify AI-generated output.
- Security: The emergence of new security loopholes within the system.
- Debugging: Increased complexity in identifying and fixing errors.
- Cognitive Debt and AI Burnout: Over-reliance on AI leads to a loss of context for developers, a phenomenon described as "AI burnout," where the human element loses the deep understanding of the system infrastructure.
Poorly managed open-source repositories often create "AI slope," contributing to the degradation of code quality.

Frameworks and mechanisms are essential for building reliable and scalable systems. The D3 framework, developed at Amazon for agentic coding (conceptually noted in the context of 2025 developments), consists of three distinct phases:
Discover:
- Involves the creation of a "research artifact."
- This document must contain comprehensive information about the repository, including all upstream and downstream dependencies.
Plan (Define):
- Based on the research artifact, a strictly defined plan is created.
- This phase involves breaking down the requirements into many specific, well-defined tasks.
Deliver:
- The AI takes the information from the research and planning stages to execute and solve the specified problem.

Human-in-the-Loop: AI cannot be fully trusted at the current stage of development. Every process must include a human sign-off component.
Model Diversity: While solving problems, it is critical to use multiple different models. These models should not just be from the same family (e.g., different versions of the same provider’s model) but should come from entirely different companies to ensure robust verification.
Agent Constraints: Instead of swarms, systems should prioritize using a small number of well-defined agents. A recommended limit is approximately 4 specialized agents.
Efficiency Metric: The primary objective of engineering teams should shift from "token maxing" (optimizing for the volume of AI generated content) to "ROI maxing" (optimizing for the return on investment and business impact).