The era of casual, conversational AI coding is ending, replaced by a more rigorous discipline called 'agentic engineering' that emphasizes oversight, architecture, and orchestration over raw speed. In February 2025, AI researcher Andrej Karpathy introduced the term "vibe coding" to describe a new way of building software: describe what you want in plain English, let an AI generate the code, iterate through chat, and repeat. It was fast, fun, and felt revolutionary. Exactly one year later, Karpathy announced that phase is effectively over.

What Changed Between Vibe Coding and Agentic Engineering?

The shift isn't about abandoning AI-assisted development. It's about maturing it. Karpathy proposed a new framework called "agentic engineering" to describe the next phase, where developers spend most of their time orchestrating AI agents rather than writing code directly. "'Agentic' because the new default is that you are not writing the code directly 99% of the time," Karpathy explained. "You are orchestrating agents who do and acting as oversight. 'Engineering' to emphasize that there is an art and science and expertise to it".

This distinction matters because it reframes the developer's role entirely. Instead of being a code writer, you become a director, reviewer, and architect. The leverage is real, but it requires discipline. Karpathy noted that top-tier agentic engineering requires "high-level direction, judgment, taste". Those three qualities separate production-grade AI-assisted development from demos that look impressive but don't scale.

Karpathy

Boris Cherny, head of Anthropic's Claude Code, is living this transition in real time. In early 2026, Cherny revealed that 100% of his production code is now generated by AI, and he hasn't written a line of code by hand since October 2025. Speaking at Anthropic's developer conference and in interviews, Cherny made clear that he finds the phrase "vibe coding" actively counterproductive because it implies sloppiness and lack of rigor. When asked what should replace it, he reportedly asked Claude itself, which suggested "agentic engineering," echoing Karpathy's framework.

"Engineers at Anthropic are no longer writing code. They are reviewing, directing, and governing the agents that do the work," Cherny observed.

Boris Cherny, Head of Claude Code at Anthropic

How Does Spec-Driven Development Work in Practice?

The practical difference between vibe coding and agentic engineering comes down to structure. Vibe coding lives in chat windows, iterating on requests without persistent context. Agentic engineering inverts that workflow by starting with a specification document.

Mariya Mansurova, writing in Towards Data Science, demonstrated this approach by building a fully functional personal fitness tracking web application in 4.5 hours using Claude Code, guided entirely by spec-driven development principles. Before a single line of code was generated, she produced structured markdown documents stored in the repository that included:

• Architecture overview: High-level system design and component relationships
• Feature specifications: Detailed requirements for each feature phase
• API contracts: Exact input and output formats for integration points
• Testing strategy: What validation looks like before code is considered complete

Only after this architecture was validated through review and even a second-agent "critique pass" did implementation begin. Each feature then followed a disciplined cycle: plan, implement, validate. Crucially, the specification was updated continuously and served as the source of truth for both human developers and AI agents. Context doesn't decay across sessions because it's externalized into the repository, not kept alive in a chat window.

"With agentic engineering, the role of the developer shifts toward steering, reviewing, and making architectural decisions, rather than directly writing specifications or code," Mansurova noted. This shift is already standardizing across the industry. GitHub released Spec Kit, an open-source command-line tool that brings the spec-driven development workflow to any agent environment with 30 integrations already supported. JetBrains, working with DeepLearning.AI, formalized this into a short course titled "Spec-Driven Development with Coding Agents".

How Are Claude Code and Codex Competing on Benchmarks?

While the methodology is standardizing, the tools themselves are in intense competition. In May 2026, Claude Opus 4.7 and OpenAI's GPT-5.5 (powering Codex) both shipped major upgrades, and the benchmark results show a split victory.

Claude Opus 4.7 achieved 64.3% accuracy on SWE-bench Pro, a rigorous benchmark that tests real-world software engineering tasks. GPT-5.5 leads on SWE-bench Verified with 88.7% accuracy and dominates Terminal-Bench 2.0 with 82.7% accuracy. The practical implication: Claude Code excels at complex, coordinated tasks that require deep codebase understanding, while Codex optimizes for speed and terminal-first workflows.

The scale of Claude Code's adoption is striking. Claude Code authors approximately 326,000 GitHub commits per day, representing roughly 10% of all public commits on GitHub. This metric alone signals that agentic engineering isn't theoretical; it's already the default workflow for a significant portion of the developer community.

Both tools now support multi-agent workflows. Codex shipped subagents to general availability on March 14, 2026, allowing up to 8 parallel agents to work simultaneously in isolated cloud sandboxes. Claude Code's Agent Teams use a coordinated approach where sub-agents share task lists with dependency tracking and can message each other directly. For independent tasks, Codex's isolation model wins on speed. For complex refactors with dependencies between subtasks, Claude's coordinated teams provide better orchestration.

What Does This Mean for Managed Service Providers and Teams?

The shift from vibe coding to agentic engineering creates a significant market opportunity. In April 2026, KPMG published a managed services outlook survey finding that more than 90% of executives now view managed services as essential for delivering agentic AI at scale. The gap between wanting agentic AI capabilities and being able to build and govern them internally is enormous, and that gap is where managed service providers traditionally operate.

However, the managed services market for agentic engineering differs fundamentally from earlier generations of AI tooling. Customers moving from vibe coding to agentic engineering aren't looking for someone to provision compute or manage service-level agreements. They're looking for partners who can establish governance frameworks, review agent-generated code, and architect spec-driven development workflows.

For individual developers and small teams, the practical implication is clear: the tools have matured enough that you can now build production software with AI agents, but only if you treat it like engineering, not like a chatbot. That means writing specifications before code, maintaining version control discipline, automating tests, and reviewing agent outputs with the same rigor you'd apply to human code reviews. The leverage is real, but the responsibility is real too.