Vibecoding? Agentic engineering!

What is Agentic Engineering?

Behind the hip term "vibecoding" lies what may prove to be the most transformative innovation brought to us by artificial intelligence: the emergence of Agentic Engineering.

The casual moniker "vibecoding" obscures the profound shift occurring in how complex problems are solved. Consider two fundamental observations that reveal the true nature of this transformation:

First, even a few months of practice demonstrates that successful "vibecoding" depends not on following ephemeral "vibes," but on adhering to meticulous rules and well-defined processes. The practitioners who achieve consistent quality results are those who have developed systematic approaches -- understanding when to use plan mode versus immediate execution, how to structure context effectively, when to employ sub-agents for parallelization, and how to iteratively refine specifications before implementation.

Second, while current implementations focus predominantly on software development, the underlying mechanism -- LLM-driven analysis, planning, and execution -- is fundamentally domain-agnostic. Today's "vibecoding" represents merely the first application of a broader capability: AI agents that can understand requirements, decompose complex problems, generate detailed implementation plans, execute those plans through appropriate tools, and iterate based on feedback. This same capability can be applied to electrical engineering, mechanical design, business process optimization, content creation, and countless other domains requiring systematic problem-solving.

Agentic Engineering is the discipline of developing engineering solutions by defining requirements, directing execution, and evaluating outputs of AI agents that autonomously decompose tasks, generate implementations, and iterate toward specifications.

Agentic Engineering as the democratization of engineering

Every conscious observer recognizes that our world remains profoundly "under-engineered." For every problem that receives an engineering solution, hundreds stay unaddressed -- not because solutions are impossible, but because engineering capacity has always been scarce and expensive.

Throughout history, a barrier has stood between problems requiring engineering solutions and their actual implementation: the cost of engineering effort. This barrier shaped civilizations. Florence and Milan could erect magnificent cathedrals not because other cities lacked ambition or need for grand religious architecture, but because only these extraordinarily wealthy city-states could afford the engineering teams capable of designing and constructing such monuments. The engineering capacity existed in theory for hundreds of cities; the economic reality permitted it in only a handful.

This pattern has persisted across centuries and scales. When solving a problem requires assembling a team of qualified engineers, organizations must first accumulate substantial capital to compensate those engineers. Even in our contemporary economy, with dramatically expanded engineering education and tools, the barrier remains substantial. Engineering talent is perpetually scarce relative to the problems awaiting solutions.

Consider a prosaic but ubiquitous example: a mid-sized trading company seeking custom software to manage its operations more efficiently. There are hundreds of thousands of such companies globally. Until recently, each faced stark choices: hire and maintain a team of 3-4 software engineers (system architect, UI designer, fron-end developer, back-end developer, etc.) that requires significant ongoing capital, or outsource development that is expensive and often misaligned with specific needs. The result? Most businesses left the majority of their potential engineering improvements unaddressed. Features that would save hours weekly remained unbuilt. Integrations that would eliminate manual data entry remained unconstructed. Custom analytics that would illuminate business patterns remained unconceived. Not because these improvements lacked value, but because the barrier to engineering implementation exceeded the perceived return.

The mathematics were brutal but simple: if engineering solutions require teams of expensive human specialists, and there exist far more problems than available engineering teams, then most problems will remain unsolved indefinitely.

Agentic Engineering fundamentally alters this equation. When AI agents become capable of understanding requirements, generating implementation plans, writing and testing code, and iterating based on feedback all at marginal cost approaching zero, the economic barrier to engineering solutions effectively vanishes.

The key insight is not merely that AI makes engineering faster or cheaper -- though it does both. The transformative effect is that AI agents make engineering abundant where it was scarce, and accessible where it was exclusive. When the barrier to solving a problem through engineering approaches zero, the number of problems that will be solved through engineering increases thousands of times.

The implications cascade across every sector:

• Businesses: Companies of any size can now engineer custom solutions for their specific workflows, rather than forcing their processes into generic software templates.
• Municipalities: Even small towns can implement sophisticated AI information services, automated permit systems, infrastructure monitoring, and citizen engagement platforms -- capabilities previously available only to major cities with large technology budgets.
• Healthcare: Individual practices and small hospitals can develop 24/7 response systems, patient monitoring platforms, and specialized diagnostic tools tailored to their specific patient populations and clinical workflows.
• Education: Individual schools can create custom learning platforms adapted to their pedagogy, student population, and available resources -- not just selecting from limited commercial offerings.
• Research: Scientists can rapidly develop specialized data analysis tools, simulation environments, and experimental control systems without waiting for centralized software development resources.

Consider the implications: if engineering implementation cost drops by 100x, the number of problems that become economically viable to solve through engineering increases by perhaps 1000x. Each small business that previously couldn't justify custom software can now engineer solutions. Each municipal department that made do with generic tools can now build purpose-fit systems. Each research lab that accepted manual processes can now automate and optimize.

A seductive but false narrative has emerged claiming that domain experts will simply engineer solutions for themselves—that doctors will build diagnostic systems, lawyers will create case management tools, teachers will develop learning platforms. This narrative dangerously underestimates what engineering demands.

A cardiologist possesses deep expertise in cardiac pathology but lacks training in systematic decomposition, abstraction, specification precision, and iterative refinement. A trial lawyer masters legal strategy but hasn't developed the habits of thinking in systems, evaluating technical quality, or debugging failed implementations. Domain expertise and engineering capability are fundamentally different cognitive skillsets.

AI agents don't eliminate the need for engineers. Every one of these newly viable engineering projects still requires someone who can:

• Articulate requirements with precision
• Decompose complex problems into addressable components
• Evaluate whether proposed solutions actually work well
• Iterate systematically toward quality outcomes
• Make architectural decisions about system design

The economics are stark: we're about to need perhaps 1000x more people capable of engineering thinking than we needed when implementation cost created the primary bottleneck. Not 1000x more pure programmers (AI agents increasingly handle that), but 1000x more people who can think like engineers while directing AI agents as implementation partners.

World would be transitioning from a situation when engineering capacity is the limiting factor in problem-solving, to a moment when every task that can be solved by engineering would be solved by engineering. Needless to say, this will expand a number of engineering job exponentially, and this is not hyperbole. It is the logical consequence of making engineering abundance as plentiful as computational resources became after the cloud revolution, or as information became after the internet.

Vibecoding as the canary in the coal mine

The transformation from theoretical possibility to practical reality is best understood through the lens of vibecoding and the emerging practice revealed in a recent LinkedIn Learning panel with James Montemagno (Microsoft), Ado Kukic (Anthropic), Kat Kampf (Google), and Piers Bogan (VS Code lead, Microsoft). What began as an experiment in terminal-based AI interaction has evolved into

The panel discussions have revealed that practitioners are developing three fundamental skills of agentic engineering:

(1) Articulating requirements with precision

Kat Kampf describes starting "not in a CLI or an IDE... I'll start talking to Gemini... talk through ideas and then kind of get to a PRD place." She then has "the AI go research it and tell me what competitors are out there, what's the landscape look like" - externalizing product research before specification.

Piers Bogan scaffolds projects manually first: "One thing I still feel like AI is not amazing at is like scaffolding. So like, I'll... scaffold it, then let the AI work." He understands what the agent can handle autonomously versus what requires human architectural decisions upfront.

(2) Directing execution and making architectural decisions

The most revealing pattern: Bogan creates three different coding agents to implement the same feature, runs them in parallel, then synthesizes the best elements. "Why would you ever create three versions of the thing right before? But like, that's something that's now possible with AI... because it's non-deterministic, you also get three different outputs."

This is agentic engineering in action - not accepting the first solution, but orchestrating multiple autonomous attempts and evaluating quality across variations.

James Montemagno goes further: "I said, I am not going to write a single line of code. I refuse... I only have so many key strokes left." By forcing himself to work entirely through direction rather than implementation, he developed fluency in the actual skill: "I don't even run the code locally anymore. I don't need to because I'm in that flow."

(3) Evaluating outputs and iterating systematically

When asked what defines a top engineer in the AI era, Bogan emphasizes: "You're being up-leveled with AI, right? I have to spend less time remembering the exact syntax... but I know what good looks like as an engineer."

The skill isn't writing code - it's recognizing quality, making architectural tradeoffs, and iterating toward better solutions. Montemagno argues that "AI fluency in development is going to be super important... mandatory." Interviews will ask: "How are you prompting? Why did you make that selection?... Why did you use this model over that model?"

The learning curve nobody expected

Heavy Anthropic users spend over $1,000 monthly on Claude Code - not from waste, but from discovering that mastery enables solving more problems. As engineers become more skilled at working with AI agents, they discover more experiments they can run, more variations they can explore. The barrier isn't cost - it's "imagination and skill in formulating problems effectively."

This parallels exactly what agentic engineering predicts: when implementation cost approaches zero, the bottleneck becomes the capacity to think systematically about problems and evaluate solutions. The number of people who can do this work becomes the limiting factor.

Ado Kukic identifies applications beyond greenfield development: "There is a ton of use cases for long running agents for things like migrations. If you have a piece of software that is aging and you want to modernize it... point the agent at it and say, hey, this is using C# version seven, let's bump it up to the latest."

The underlying mechanism - understanding requirements, decomposing problems, generating implementation plans, executing through tools, iterating based on feedback - applies equally to code migrations, test coverage generation, or feature development. Kukic envisions extending this to "building full production-ready applications in a single prompt."

The present as prologue

The panel operates with striking humility about what they've built. When asked about future form factors, Bogan admits: "No one knows what these form factors are... I think our team is in experimentation mode."

Yet the data point matters: 70-80% of Anthropic's technical employees use Claude Code daily. When the people building frontier AI systems choose to work this way, it signals something profound about where engineering is headed.

What vibecoding demonstrates isn't the final form of agentic engineering, but proof that the concept works at scale. Software development is merely the first domain - the easiest case because code is already formal and testable. If agentic approaches transform how professionals solve systematic problems in coding, the same mechanisms will reshape problem-solving across every domain admitting of engineering solutions.

The casual observers see "vibecoding" and imagine it's about prompts and vibes. The practitioners understand they're developing an entirely new discipline: the systematic practice of directing autonomous agents to solve complex problems - which is precisely what agentic engineering is.

Early research: what science says

In the research paper "Generative AI and Agentic Architecture in Engineering and Manufacturing: Potentials and Practice of Scalable AI Solutions" by Larichev, Masek, Chouhan, and Spiess (2025), the authors demonstrate that GenAI can increase available resources in research and development by up to 36%, directly supporting the argument that agentic engineering fundamentally alters the economics of problem-solving.

The paper describes a practical architecture combining a central orchestrator with specialized agents that handle specific tasks across the engineering value chain, validating the claim that successful implementation depends on systematic approaches rather than informal experimentation. Authors explicitly confirm that the underlying mechanism of LLM-driven analysis, planning, and execution is "fundamentally domain-agnostic," corroborating the position that today's software-focused applications represent merely the first instance of a broader capability applicable to electrical engineering, mechanical design, and business process optimization.

However, the authors reveal a significant implementation gap: while 95% of surveyed companies believe AI will play a crucial role, only 12% can effectively integrate and scale AI solutions, with over 60% still struggling in the experimentation phase.

This data underscores the critical need for engineering education transformation, as companies face challenges including data fragmentation, lack of interoperability, and insufficient training—precisely the barriers that necessitate developing engineering thinking capabilities in a vastly larger population than traditional engineering programs currently serve.

Another research paper: "Agentic Software Engineering: Foundational Pillars and a Research Roadmap" by Hassan et al. (2025), argues that the field has moved beyond AI-augmented development (SE 2.0) into a new era of Agentic Software Engineering (SE 3.0) where intelligent agents achieve complex, goal-oriented objectives rather than simply generating code.

The paper introduces the Structured Agentic Software Engineering (SASE) framework, built on a fundamental duality: SE for Humans (SE4H), which redefines the human role as an "Agent Coach" focused on high-level intent and mentorship, and SE for Agents (SE4A), which establishes structured environments where multiple agents operate effectively. Supporting this framework, the authors present evidence that practitioners achieving 100x-1000x productivity have mastered nascent best practices of directing agent teams, demonstrating that "the core skill of the modern super software engineer is mastering the duality of control: strategically deciding when to impose a structured workflow and when to 'let the agent loose.'"

The paper validates that autonomous coding agents like Claude Code, Google's Jules, and Anthropic's Devin are already responsible for hundreds of thousands of merged pull requests, with 70-80% of Anthropic's technical employees using Claude Code daily.

Authors emphasize that vibecoding and similar practices represent "the first widely-adopted implementation of true agentic engineering," serving as proof that collaboration with AI agents as partners—rather than using them as passive tools—fundamentally transforms software development from individual coding into orchestrated, team-scale engineering.

A third paper that I have reviewed was the research paper "Intelligent Design 4.0: Paradigm Evolution Toward the Agentic AI Era" by Jiang et al. (2025). In this paper, the authors propose that engineering design is entering a transformative fourth stage driven by foundation model-based multi-agent AI systems, which they term Intelligent Design 4.0 (ID 4.0).

The paper traces the evolution of intelligent design through four distinct stages: rule-based expert systems (ID 1.0), machine learning-driven design (ID 2.0), foundation model-empowered design (ID 3.0), and the emerging paradigm of multi-agent collaboration (ID 4.0). Supporting this framework, the authors present an ontological structure mapping the traditional five-stage engineering design workflow -- Task Clarification, Concept Design, Embodiment Design, Detailed Design, and Design Optimization -- onto specialized AI agents including the Requirement Analysis Agent, Concept Generation Agent, Embodiment Generation Agent, Detailed Modeling Agent, and Design Optimization Agent.

The paper demonstrates that while single foundation models like LLMs have shown promise in supporting isolated design tasks, they remain insufficient for handling the full complexity of end-to-end engineering design, necessitating coordinated multi-agent systems where specialized agents collaborate through structured workflows, shared memory systems, and dynamic task allocation.

The authors emphasize that ID 4.0 represents a fundamental shift from human-centric design implementation toward human-agent partnership, where humans act as "Agent Coaches" who orchestrate, mentor, and provide strategic oversight while autonomous agents handle execution, paralleling the exact transformation described in vibecoding where practitioners transition from writing code to directing fleets of AI agents through specification, orchestration, and evaluation.

Unique role and opportunity for students and educators

Historically, we needed far fewer people with engineering mindsets than we needed engineering solutions. Society developed elaborate systems to route problems to the small number of people who could think systematically about solutions. A business owner didn't need to think like an engineer—they needed capital to hire engineers. A hospital administrator didn't need engineering capabilities—they needed budget to purchase engineered systems.

This arrangement worked adequately when engineering capacity was scarce and expensive anyway. Why worry about the limited number of people capable of engineering thinking when you couldn't afford to employ them even if they existed?

But now we need to develop “new engineering” capabilities in perhaps 1000x more people than we were educating as engineers before.

Agentic engineering DOES NOT eliminate the need for engineering education. AI tools won’t make everyone an engineer automatically. This narrative ignores everything we've learned: successful agentic engineering requires systematic thinking, quality evaluation, iterative refinement, and architectural judgment. These capabilities don't emerge spontaneously from domain expertise.

At the same we must acknowledge that traditional engineering curricula must be 100% re-imagined. When AI agents handle implementation, the value of technical skills diminishes while the value of engineering thinking (abstraction, decomposition, specification, evaluation) increases dramatically alongwith value of managerial skills, ability to orchestrate autonomous agentic work.

Academia faces several urgent tasks:

1. Identify the core competencies: What specific knowledge and skills must the "new engineers" possess? Which aspects of traditional engineering education remain essential? Which can be deprioritized as AI agents increasingly handle implementation?

2. Develop pedagogical approaches: How do we teach engineering thinking to a much broader population than traditionally pursued engineering degrees? What scaffolding helps students develop systematic thinking habits?

3. Create practical curricula: What does a program look like that develops both engineering mindset and AI collaboration capabilities? How do we balance theory, practice, and tool proficiency?

4. Scale effectively: How can institutions reach 10x or 100x more students than current engineering programs serve, without proportionally expanding faculty or diluting quality?

The stakes are considerable. Educational institutions that successfully develop these new curricula will prepare graduates for a world where engineering capacity is abundant, where the ability to think like an engineer and direct AI agents becomes the most valuable professional capability across virtually every domain.