AGENTIC ENGINEERING
Agentic Engineering: What Comes After Vibe Coding
On vibe coding's first birthday, Andrej Karpathy renamed the professional version of it agentic engineering: agents write the code, you orchestrate them and own the oversight. The leverage is real — 46% of new GitHub code is already AI-generated — but only if the system around the agents is engineered. VibeFuse is a free, working harness for that system on Windows.
What is agentic engineering?
Vibe coding — the term Karpathy coined on February 2, 2025 — meant giving in to the vibes, accepting generated code, and forgetting the code even exists. It was perfect for prototypes and weekend projects. Twelve months later the same workflow had become the default for professionals, with one difference: oversight. In February 2026, Karpathy named that professionalized workflow agentic engineering, writing that programming via LLM agents is now “a default workflow for professionals, except with more oversight and scrutiny” and that the goal is “to claim the leverage from the use of agents but without any compromise on the quality of the software.”
His breakdown of the name is the job description:
- Agentic — “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 — “there is an art & science and expertise to it. It's something you can learn and become better at.”
So agentic engineering is not a new tool category — it is the discipline of running AI agents as a production system: owning the spec, designing what context and tools the agents get, gating their output with verification, and staying the human in the loop where decisions actually matter.
Why the name changed: the 2026 numbers
The rename is a response to data, not aesthetics. Adoption exploded while trust in the output collapsed — the gap that separates casual vibing from engineered agent workflows.
| Signal | 2026 data |
|---|---|
| US developers using AI coding tools daily | 92% (industry survey aggregation) |
| New GitHub code that is AI-generated | 46% (GitHub platform data) |
| Developers who trust AI code accuracy | 29% — down from 40% in 2024 |
| Major issues in AI co-authored PRs | 1.7x more than human-written (CodeRabbit review data) |
| AI-generated samples with OWASP Top-10 flaws | 45% (Veracode controlled testing) |
| Vibe coding users who are non-developers | 63% (Vercel) |
| Vibe coding tools market | $4.7B in 2026, ~38% CAGR |
Analysts converged on the same reading. Forrester's Andrew Cornwall: vibe coding is “great for prototypes but not wonderful for brownfield or production code.” Caylent's CTO Randy Hunt went further: “the differentiator isn't which LLM you picked, it's the agentic harness… If you don't engineer the harness, you don't get compounding leverage; you get compounding cognitive debt.”
The four practices of agentic engineering
1. Own the spec, not the keystrokes
You design the spec with your agent before code exists — the docs, the invariants, the API decisions. Karpathy: you are in charge of the taste, the engineering, the design, and the agents fill in the blanks.
2. Verification gates
Tests run before you look at code. Diffs reviewed while context is on screen. Security loops on anything touching auth, secrets, or user data — 45% OWASP failure rates do not fix themselves.
3. Harness engineering
The same model scores 35% vs 50% depending on the harness around it. Context management, tool access, and failure recovery are engineering decisions — see harness engineering.
4. Orchestration
Multiple agents on parallel tasks, a terminal and browser beside them, one shared canvas. 2026 teams juggle four AI tools on average — orchestration is how they compound instead of collide.
Vibe coding vs agentic engineering
| Vibe coding (2025) | Agentic engineering (2026) | |
|---|---|---|
| Who writes the code | You prompt, AI suggests, you accept | Agents write it; you orchestrate and review |
| Posture toward the code | “Forget the code exists” | Own the spec, review every diff |
| Best fit | Prototypes, demos, weekend projects | Production software with oversight |
| Quality mechanism | Iterate until it seems to work | Verification gates, tests, security loop |
| Failure mode | Shipping code nobody understands | Cognitive debt if the harness is not engineered |
Run your agentic engineering loop free
VibeFuse is the first ever free widget-based AI harness: Claude Code, Codex, Gemini, Cursor Agent, and Qwen run as live widgets beside terminals and browsers on one Windows canvas. Named sessions persist your loop design, MCP tools hand agents real capabilities, local Whisper and Piper keep voice on-device, and the open-source marketplace pays creators an 80% revenue share for the widgets and skills you build. The harness costs $0 — you pay only the model vendors you already use.
Practical next steps: see how it compares as a Cursor alternative or a Claude Code GUI, learn the security review loop every agentic workflow needs, and browse the 2026 vibe coding statistics behind the rename.
- ✓ Multi-agent canvas
- ✓ MCP + Skill Seekers
- ✓ Local Whisper + Piper
- ✓ Named sessions
- ✓ 80% creator payouts
Explore VibeFuse & harness guides
- Harness Guide
- Free Coding Tools
- AI Coding Agent Desktop
- Free Voice Transcription
- Free Text to Speech
- VibeFuse Product
- Widget Marketplace
- Download Free
- VibeFuse Docs
- Shareable AI Widgets
- Shareable AI Skills
- MCP Tools
- AI Agent Harness
- Harness Engineering
- HyperFrames Video
- Skill Seekers
- Sell AI Skills
- AI Skills Marketplace Compared
- Cursor Alternative
- Local Whisper + Piper
- Custom AI Dashboard
- Community Hub
Agentic engineering FAQ
What is agentic engineering?
Agentic engineering is the term Andrej Karpathy proposed in February 2026 for the professionalized version of vibe coding: agents write the code while you orchestrate them, own the spec, and act as oversight. The "engineering" part is deliberate — there is an art, science, and expertise to directing agents, and it is a skill you can learn and improve.
Why did Karpathy rename vibe coding?
Vibe coding started as a hands-off style for throwaway projects — "forget that the code even exists." A year later, programming via LLM agents had become the default professional workflow, but with more oversight and scrutiny. Karpathy said the goal now is to "claim the leverage from the use of agents but without any compromise on the quality of the software," so the casual name no longer fit.
What is the difference between vibe coding and agentic engineering?
Vibe coding is prompt-driven and acceptance-based: you describe, the AI generates, you iterate until it seems to work — fine for prototypes. Agentic engineering treats agents as a production system: you design the spec, engineer the harness (context, tools, recovery), gate output with tests and security review, and stay accountable for what ships. Same leverage, different posture toward quality.
What does an agentic engineer actually do?
Four things: own the specification (the docs, invariants, and design decisions agents fill in around), engineer the harness that runs the agents (context management, tool access, failure recovery), run verification gates (tests, diffs, security loops on auth and secrets), and orchestrate multiple agents in parallel without letting their work collide. You write less code and make more decisions.
Do I need a harness for agentic engineering?
The evidence says yes: identical model weights score 35.2% vs 49.6% on Terminal-Bench depending on the harness they run in. Analysts at Caylent put it plainly — the differentiator is not which LLM you picked, it is the agentic harness; without engineering it you accumulate cognitive debt instead of compounding leverage. VibeFuse is a free working harness for Windows: five agent CLIs as widgets, MCP tools, sessions, and a marketplace.
Is agentic engineering just a rebrand of prompt engineering?
No. Prompt engineering optimizes single requests to a model. Agentic engineering runs a system: persistent agents that read your repo, run commands, edit many files, and fail and recover — with you supplying the spec, the harness design, and the quality gates. It is closer to what an engineering manager plus tech lead does, translated to directing AI agents.