What is agentic engineering?

Agentic engineering emphasizes agentic programming as a tool rather than the force building the entire codebase end-to-end. The term builds upon the term  “vibe coding“ coined by OpenAI cofounder Andrej Karpathy in 2025. Vibe coding describes the free-form practice of prompting AI tools to generate code rather than writing code manually.

However, with the start of 2026, the software industry needs terminology that better reflects professional development practices. Hence, Karpathy coined a new term: “agentic engineering.” In this article, we will explore what agentic engineering means and how it differs from vibe coding.

With the rapid growth of artificial intelligence (AI) and AI agents, using generative AI for building software has become a default part of the standard software development workflow. In Stack Overflow’s 2025 Developer Survey, it was found that 84% of respondents use or intend to use AI-assisted programming in their development process.

However, the survey also shows more developers distrust the accuracy of AI tools than trust them, with 46% expressing skepticism compared to 33% who feel confident in the results.

Only a small share (3%) says they “highly trust” AI‑generated output. Seasoned developers are especially wary, reporting the lowest rate of “highly trust” (2.6%) and the highest rate of “highly distrust” (20%).¹ This pattern highlights a broad expectation that human oversight remains essential, particularly for roles with significant responsibility.

In practice, developers use AI for tasks that are time-consuming but relatively low risk when guided by an experienced human reviewer. This action includes tasks such as code refactoring, generating boilerplate code and tests, performing lightweight code reviews, drafting documentation, scaffolding APIs and other low risk tasks.

To recap, vibe coding allows developers to automate and prototype quickly with a few natural language prompts using large language models (LLMs) or agentic systems.

The word “vibe” in the term “vibe coding” has a casual, improvisational connotation. While this framing was well‑suited to the early, exploratory phase of AI‑assisted programming, it clashes with the current reality of developers’ skepticism toward AI coding agents.

Without the proper expertise in using LLMs for software engineering purposes, vibe coding can produce what is called “AI slop,” code that is not useful or breaks existing code. Such broken code often increases the technical debt of engineering teams with much of their time spent understanding, debugging and refactoring the code.

Alternatively, the term “agentic engineering” is detailed by Karpathy on X in a different way:

• “Agentic”: An orchestration of agents writes the code, and the human developer oversees and validates the output. As the agent or multi-agent system iterates through the subtasks, we maintain a human-in-the-loop.
• “Engineering”: There is a level of expertise required to use agentic workflows for meaningful code production that does not jeopardize the code quality. This skill is one that can be developed and improved.

Want to see how agentic engineering changes the role of the software engineer? Watch Agentic Engineering vs. Software Engineering: Beyond Vibe Coding to explore how agentic engineering differs from traditional software development and how developers can work alongside AI agents.

Agentic engineering works by combining human engineering expertise with AI agents that can perform specific software development tasks. Instead of asking an AI system to generate an entire application from a single prompt, developers define goals, break complex work into manageable tasks, provide agents with the necessary context and tools, and review and validate the resulting work.

A typical agentic engineering workflow includes the following steps:

1. Define the goal and requirements. Developers establish what the software or feature needs to accomplish, along with technical requirements, constraints and quality standards.

2. Break the work into tasks. Complex development work is divided into smaller, well-defined tasks that an AI agent can execute, such as writing code, updating documentation, generating tests or identifying bugs.

3. Assign tasks to AI agents. Developers select or configure agents for specific tasks and give them access to the relevant code, documentation, development tools and other context they need to work effectively.

4. Orchestrate and guide the agents. Developers provide instructions, establish dependencies between tasks and determine when agents should work independently or when their output requires human intervention.

5. Review and validate the output. Human engineers evaluate the agent-generated code and other outputs for correctness, security, performance, maintainability and alignment with requirements. Automated tests and other evaluation tools can support this process.

6. Iterate and refine. Agents can revise their work based on test results, developer feedback or changing requirements. Developers continue to guide the process until the output meets the required standards.

7. Integrate approved changes. Once the work has been reviewed and validated, developers integrate the resulting code into the broader software development workflow and production codebase.

The key distinction is that AI agents perform much of the execution, while human engineers remain responsible for directing the work, making high-level technical decisions and ensuring the quality of the final result. This allows organizations to use AI for more than code generation while retaining the engineering judgment and oversight required for professional software development.

Organizations can adopt agentic engineering by establishing the right governance, equipping developers with the skills to work with AI agents and creating the technical infrastructure needed to support agentic workflows. This includes providing agents with relevant organizational context, breaking development work into manageable tasks and establishing clear processes for human oversight and validation.

Organizations can take several steps to support adoption.

Together, these practices can help organizations move from experimentation with AI coding tools toward a more structured approach to agentic engineering, where AI agents take on increasingly complex development tasks under appropriate human oversight. 

Agentic engineering allows us to shift away from deterministic logic toward the nuanced territory of probabilistic judgment. Hence, it requires a fundamental shift in how organizations think and how engineering teams operate.

As the landscape of AI‑assisted software development continues to evolve, its terminology will likely continue to shift as well. While “agentic engineering” provides a more mature and accurate framing for how developers collaborate with autonomous systems today, the practices behind it will continue to advance as AI agents become more capable. In the coming years, we can expect agentic systems to handle increasingly complex tasks.

At the same time, the role of developers will evolve from simply writing code to designing, supervising and shaping the behavior of these AI systems. Whether you are an AI engineer, full‑stack developer, data scientist or someone beginning your coding journey, the core principles or human oversight, system design literacy and high‑judgment decision making, will remain foundational.

If you’re feeling unsure about where to begin, start small. Explore one agentic workflow, whether open source or enterprise-grade, test it with your team and see what’s possible.