Andrej Karpathy might be best known as a co-founder of OpenAI and former head of AI at Tesla, but he has also unexpectedly become a prolific neologist.
Almost exactly one year ago, he made a post on X that coined the term “vibe coding” to describe the free-form, prompt-driven way of generating software with AI that was freshly made possible. It went immediately viral. For people who could suddenly use plain language to prompt the creation and iteration of a working piece of software, riding on vibes resonated. However, vibe coding came with considerable risks.
What is vibe coding?
Read about AI developmentA year later, AI-assisted development has evolved considerably. It’s more structured, more intentional, and more professional. It has developed into an entirely different process. Karpathy now argues that the industry has moved beyond “vibes” toward something more structured and professional: agentic engineering.
“Today,” he wrote, “…programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny.”
This shift matters not just to developers experimenting with AI, but to CTOs, software leaders, and enterprise teams grappling with the risk and reward of adding autonomous AI components into their development lifecycles.
“Many people have tried to come up with a better name for this to differentiate it from vibe coding, personally my current favorite "agentic engineering": "agentic" because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight. "engineering" to emphasize that there is an art & science and expertise to it. It's something you can learn and become better at, with its own depth of a different kind.”
Andrej Karpathy
Co-Founder, OpenAI
What is agentic engineering?
Agentic engineering, also known as AI agentic programming, is a software development discipline in which humans define goals, constraints, and quality standards while AI agents autonomously plan, write, test, and evolve code under structured human oversight.
Karpathy breaks down the term specifically, writing "'agentic’ because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight. ‘engineering’ to emphasize that there is an art & science and expertise to it. It's something you can learn and become better at, with its own depth of a different kind.”
Where vibe coding was about expressing intent and watching tools spin up code quickly, agentic engineering recognizes that professional software development requires planning, verification, iterative refinement, and governance to be secure, scalable, and viable for real business use. This more accurately reflects how professional engineering teams collaborate with AI today.
Agentic engineering is based on delegation and supervision. Instead of treating an LLM as a glorified autocomplete, developers define goals, constraints, quality criteria, and workflows for the AI. A suite of AI agents, each with specialized capabilities, then works toward those high-level targets, generating, testing, debugging, and integrating components under human direction.
Want to know more about agentic engineering terms?
Read the GlossaryThe benefits of agentic engineering are significant. By approaching AI development in a more structured way, engineering teams can reduce variability in quality, shorten feedback loops between idea and implementation, lower the cost of maintaining complex systems, and make software development more predictable.
How is agentic engineering different than vibe coding?
Early vibe coding posed significant risks, making enterprises and large teams wary of adopting it. Vibe-coded software often included major security vulnerabilities. It was also difficult to audit, leading to technical debt and brittle systems that were hard to maintain.
Agentic engineering tackles these concerns explicitly:
- Structured Oversight: Instead of letting an LLM create unreviewed code, agentic workflows embed quality gates, automated testing, and audit trails throughout. AI agents generate artifacts, but humans evaluate and validate them before they are merged into production systems.
- Goal-Driven Decomposition: Agents are given defined objectives rather than open-ended prompts. This prevents aimless code generation and focuses output on verifiable outcomes aligned with business requirements.
- Iterative Verification: Multiple agents can operate together, for example, one focused on implementation, another on testing, and yet another on security review. This multi-agent coordination mirrors established engineering practices while benefiting from the speed of AI automation.
- Governance and Traceability: Agentic engineering workflows enforce traceability from requirement to code output, enabling teams to audit decisions, verify results, and meet compliance needs.
This doesn’t mean that vibe coding will disappear. The opposite is true. There are still many great uses for vibe coding, especially in prototyping, developing MVPs, and ideation and experimentation. Vibe coding lets a single person quickly launch a project or test an idea. Agentic engineering is developing as a separate discipline, oriented more towards professional engineering teams working on critical business projects.
How are engineering teams using AI development today?
Many businesses and organizations today are actively experimenting with autonomous elements in development workflows. Tools like GitHub Copilot, Claude Code, and Cursor have extended beyond simple autocomplete to more autonomous task handling. Some platforms now let developers assign AI agents to manage issue triage, automate pull request suggestions, or even refactor code across entire modules.
Modern AI coding agents can now:
- Break down a high-level goal into subtasks
- Decide which files, APIs, or tools to touch
- Run tests or scripts autonomously
- Observe failures and retry with adjustments
For enterprises, the agentic engineering is more than a technological change; it’s organizational. Businesses are beginning to embed AI agents into core engineering workflows in ways that impact strategy, speed, and quality. While the early hype around AI coding focused on individual productivity: faster typing, fewer bugs, more output per engineer, agentic engineering focuses on systems more than individuals.
What are the business use cases for agentic engineering?
So, how will agentic engineering actually show up in businesses? These are a few of the use cases and business benefits we may see becoming standard in 2026.
- Automating routine development: AI agents will be deployed to handle repetitive engineering tasks such as stub generation, boilerplate creation, refactoring, or regression testing, freeing human engineers to focus on design, architecture, and novel problem-solving.
- Domain-Specific Agents: Teams can construct specialized AI agents for particular business functions, e.g., compliance checking in finance apps, data pipeline orchestration in analytics, or API integration verification for microservices.
- Continuous modernization: Many enterprises sit on decades of legacy code. Agentic engineering enables gradual modernization: agents can refactor components, improve test coverage, and document systems incrementally without large, risky rewrites.
- Operational resilience: Because agentic workflows are explicit and repeatable, businesses gain resilience. If a key engineer leaves, the process remains in place. Knowledge lives in orchestrated systems, not just in people’s heads.
- Compliance-driven development: In regulated industries, agentic systems will continuously enforce compliance rules during development rather than relying on end-of-cycle audits. For example, an agent can block changes that violate data residency rules or security policies before they ever reach production. This can be essential for sectors like healthcare, finance, and critical infrastructure.
- Internal business software: Companies will increasingly use agentic engineering to build and maintain internal tools like dashboards, approval systems, inventory apps, and finance workflows. Agents can implement incremental updates faster than traditional teams while preserving consistency, and no-code platforms with agentic abilities provide a stable platform for building and iterating on custom internal tooling.
Learn how to develop business apps with AI
Read moreEmbracing AI development as a business function
The AI-software ecosystem is rapidly maturing. What began as exploratory experiments with generative models has grown into a structured, strategic approach to building software with autonomous agents. The role of developers is inherently going to need to adapt to this new technological landscape. As Karpathy notes, “you are not writing the code directly 99% of the time… you are orchestrating agents who do and acting as oversight.”
What we are seeing is the professionalization of AI-assisted software development. Where vibe coding captured the excitement of early generative tools, agentic engineering represents a more grounded reality for businesses. For engineers, this means a shift in skill sets. For businesses, it means a new way to think about software as an adaptive, semi-autonomous capability. One that can be scaled responsibly rather than improvised.
