Where did the term 'Agentic Engineering' come from?
The Rise of “Agentic Engineering”
The term “Agentic Engineering” was primarily popularized and brought into the mainstream by Andrej Karpathy, the renowned AI expert, former Tesla Director of AI, and co-founder of OpenAI.
He formally introduced this term in early February 2026 to define the next phase of AI-assisted development.
1. The Origins and Evolution
-
The Visionary: Andrej Karpathy
Karpathy previously coined the term “Vibe Coding,” which referred to humans writing code by simply describing a “vibe” to an AI. One year later, he argued that “Vibe Coding” no longer sufficed to describe the landscape, leading to the birth of “Agentic Engineering.”
-
Why “Agentic”? Because 99% of code is no longer written directly by humans; it is autonomously generated, tested, and debugged by AI Agents. Humans have transitioned into the roles of “Orchestrators” and “Supervisors.”
-
Why “Engineering”? He emphasizes that this is no longer just about tossing out prompts; it is a discipline of “art, science, and expertise.” Designing an agent’s reasoning paths, memory, and tool integration requires deep engineering foundations.
-
-
The Strategic Foundation: Andrew Ng
While the specific phrase “Agentic Engineering” is closely tied to Karpathy’s recent discourse, the concept of “Agentic Workflows” was championed by Andrew Ng as early as 2024.
-
Andrew Ng’s Contribution: He defined the four key patterns of Agentic AI: Reflection, Tool Use, Planning, and Multi-agent Collaboration.
-
The Distinction: While Ng discusses workflows from a “system design” perspective, Karpathy uses “Engineering” to formally define it as a new professional skill set and academic discipline.
-
2. Evolutionary Timeline
-
2024: Andrew Ng begins promoting “Agentic Workflows,” arguing that iterative AI reasoning is more important than raw model scale.
-
Early 2025: Karpathy introduces “Vibe Coding” to describe how non-engineers can build software.
-
February 2026: Karpathy formally proposes “Agentic Engineering,” signaling that AI-assisted development has moved from “casual experimentation” to “rigorous engineering.”
Summary: If asked who “invented” the term, the answer is Andrej Karpathy; however, the “scientific foundation” of the theory is deeply rooted in the work of Andrew Ng.
Defining Agentic Engineering
Agentic Engineering is the engineering discipline focused on designing, developing, and optimizing AI Agent systems.
Unlike traditional software engineering, which relies on predefined logic (If-Then), or standard Generative AI, which focuses on simple Q&A, Agentic Engineering builds systems with autonomy, reasoning, and execution capabilities. It enables AI to achieve complex goals without constant, step-by-step human intervention.
The Four Pillars of Agentic Engineering
To transform a Large Language Model (LLM) into a true “Agent,” four engineering dimensions must be integrated:
-
Planning: Task decomposition (breaking big goals into sub-tasks) and self-reflection (correcting errors via Chain-of-Thought).
-
Memory: Utilizing Short-term Memory (Context Window) and Long-term Memory (RAG/Vector Databases) to recall professional knowledge or past experiences.
-
Tool Use (Function Calling): Giving the AI “hands” by allowing it to call APIs, search the web, execute Python code, or manage databases.
-
Multi-Agent Orchestration: Designing a “team” of agents with specialized roles (e.g., a Coder, a Tester, and an Integrator) to complete massive projects beyond the capacity of a single model.
From “Prompt Engineering” to “Agentic Engineering”
| Feature | Prompt Engineering | Agentic Engineering |
| Focus | Optimizing the input (The Prompt) | Designing the system (The Agentic System) |
| Process | One-shot (Input->Output) | Iterative (Loop -> Test $-> Refine) |
| Complexity | Simple tasks (Summarization, Q&A) | Complex workflows (Software dev, Research) |
| Human Role | The “Writer” | The “Orchestrator” |
Why This Matters in 2026
As the reasoning capabilities of LLMs have stabilized, enterprise needs have shifted from “writing an email” to “automating a business process.” Agentic Engineering solves the issues of AI “hallucinations” and “lack of execution” by providing a closed execution environment and verification mechanisms. It transforms AI from a “talking encyclopedia” into a “highly capable virtual employee.”
