Vibe coding vs agentic coding vs AI-assisted development
Vibe coding vs agentic coding vs AI-assisted development: what each term means, where it came from, how much review it involves, and when each fits.
A PM builds a working prototype of a new onboarding flow in an afternoon, without reading a line of the code. An engineer on the same team uses Claude Code to refactor the billing module, reading every diff and running the tests. Both say they're "coding with AI." One is vibe coding. The other isn't, even though the engineer's tool is more autonomous.
The terms get mixed up constantly. This guide separates them, from the original sources, checked on 16 November 2026.
The short answer
| Vibe coding | AI-assisted development | Agentic coding | |
|---|---|---|---|
| Defined by | Not reviewing the code | Using AI help at all | An agent working in a loop |
| Who reviews the code | Nobody | The developer, as usual | Depends on how it's used |
| Typical tools | App builders like Lovable, or an agent on "accept all" | Autocomplete, chat, agents | Claude Code, Codex, Cursor's agent |
| Good for | Prototypes, personal tools, demos | Any software work | Multi-step tasks across files |
| Risk | Unmaintainable, insecure code | Normal | Rises as review drops |
The key point: the first is defined by how much you review, the third by how the AI works. They're different axes.
Vibe coding
Andrej Karpathy coined the term on 6 February 2025. He described giving in to the vibes and letting the code grow past his own understanding: accepting every change without reading diffs, pasting error messages back in, and asking for random changes when a bug wouldn't go away. His summary was that you "forget that the code even exists," and that it was fine for throwaway weekend projects.
Simon Willison, writing shortly after, pinned the definition down: building software with an LLM without reviewing the code it writes. He also warned against using the term for all AI-assisted programming.
The term spread fast. Wikipedia notes Merriam-Webster listed it as slang and trending in March 2025, and Collins named it Word of the Year for 2025.
AI-assisted development
The broad category: any software development where AI helps, from autocomplete to chat to agents. What makes it ordinary software development is that the developer still owns the result. Willison's golden rule: don't commit code you couldn't explain exactly to someone else. If you reviewed it, tested it, and can explain it, he argues, the fact that an LLM wrote it doesn't matter.
Agentic coding
Agentic coding describes how the AI works rather than how much you check it. Anthropic's working definition of an agent is an LLM autonomously using tools in a loop. A coding agent like Claude Code reads files, plans, edits several files, runs tests, reads the failures, and tries again, without you prompting each step.
That autonomy makes vibe coding easier, since an agent on "accept all" can build a lot of unreviewed code quickly. But it doesn't require it. The same agent can work in plan mode, with you approving a plan first and reviewing the diff after. Some people call that disciplined version agentic engineering.
Why the difference matters
Wikipedia's article summarizes the main criticisms of vibe coding in production:
- Nobody understands the code, so debugging and maintenance are slow.
- Security problems go unnoticed. One vibe coding platform was reported in May 2025 to have generated apps where personal data could be read by anyone, and a Veracode study found generated code improving in function much faster than in security.
- Existing systems are the hard part. Willison notes most engineering work is evolving existing systems, where understanding the code matters most.
None of these apply to a prototype that will be thrown away. All of them apply to code that ships.
A practical split for teams
- PMs and designers: vibe code prototypes and demos freely. Hand them to engineering as a reference, with a PRD, not as the code to ship.
- Engineers on production code: use agents for speed, review every diff, and keep tests as the gate.
- Internal tools with few users: a middle ground. Vibe code it, but have someone read the parts that touch data or permissions.
- Anything with customer data, money, or auth: no vibe coding.
What agents still don't know
Reviewing an agent's diff catches bad code. It doesn't catch good code built on the wrong decision: the agent cleanly implements per-seat pricing a week after the team agreed to move to usage-based pricing in a meeting it never heard about. Review checks what was written, not whether the agent knew what the team decided.
Where Lockstep fits
Lockstep closes that gap for agentic coding with Claude Code. It captures team decisions from PRDs, Slack, and pull requests, routes each to its owner for approval, and briefs each Claude Code session on the approved decisions that apply to the work. Its checks are advisory: they flag a likely conflict for a person to review.
See how it works.
Sources checked 16 November 2026
- https://simonwillison.net/2025/Mar/19/vibe-coding/
- https://en.wikipedia.org/wiki/Vibe_coding
- https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Related reading
For PMs working with agents, see how to use AI agents as a product manager. For comparing agents, see Claude Code vs Cursor vs Codex. For spec-first agentic work, see spec-driven development for product managers. For Anthropic's agent for non-code work, see Claude Code vs Cowork. For the long-term cost of unreviewed code, see technical debt from AI coding. For reviewing what agents write, see AI code review best practices. For agentic AI terms in general, see agentic AI vs AI agents vs generative AI.
Frequently asked questions
- What is vibe coding?
- Building software by describing what you want to an LLM and accepting the code it generates without reviewing it, steering by results and follow-up prompts. Andrej Karpathy coined the term in February 2025 and described it as fine for throwaway weekend projects.
- What is the difference between vibe coding and agentic coding?
- Vibe coding is about how much you review: not at all. Agentic coding is about how the AI works: an agent plans, edits files, runs commands, and checks results in a loop. You can use an agent and review everything it does, which is agentic but not vibe coding, or let it run and accept everything, which is both.
- Is vibe coding bad?
- Not for prototypes, personal tools, and experiments, where it lets non-programmers build things. It's risky for production code, because nobody understands the code well enough to maintain it, debug it, or catch security problems. Studies have found generated code improving in function faster than in security.
- Is using Claude Code or Cursor vibe coding?
- Only if you accept changes without reading them. Simon Willison's rule for responsible use is not to commit code you couldn't explain to someone else. Used that way, an AI coding agent is a tool for software development, not vibe coding.
- Can product managers vibe code?
- Yes, and it's a good use: prototypes to test an idea, internal tools, and demos for user research. Hand the idea to engineering as a prototype and a PRD rather than shipping the vibe-coded code itself.
Keep reading