Short version: AI has split software delivery into two modes. Vibe coding is fast and cheap for throwaway work, but expensive to run and risky to ship. Agentic engineering wraps the AI in specifications, tests, and human review, so the output is dependable. If you manage delivery, your estimates, budgets, risk register, and staffing plan all move.
What is vibe coding, and why should a project manager care?
Vibe coding is a way of building software where you describe what you want in plain language, accept whatever the AI produces, and paste errors back until it runs. The term was coined by Andrej Karpathy in February 2025 (karpathy on X), who described giving in to the vibes and forgetting the code even exists. It is fast and it feels like magic. It is also, by definition, code that no human has fully read.
Agentic engineering is the disciplined end of the same idea. The AI still writes the implementation, but inside human-designed specifications, automated tests, and feedback loops that check the work. These are two ends of one spectrum (Wikipedia). The differentiator is not whether your team uses AI. It is how much verification surrounds the output. Rule of thumb: a weekend prototype can be pure vibe coding, anything a customer depends on needs agentic engineering.
What changes in the software development life cycle?
The phases you know (requirements, design, build, test, review, maintain) do not disappear. Their proportions change. AI compresses implementation hard: work that took weeks can take hours. It does almost nothing for the human-paced parts, which are requirements, architecture, and verification. So the bottleneck moves. Writing code stops being the slow step. Deciding exactly what to build, and confirming it was built correctly, becomes the slow step.
This leads to what the Google whitepaper calls the factory model: the team's real output is not the code, it is the system that produces the code. A factory manager does not assemble each unit by hand, they design the line and own quality control. Requirements become a live conversation with the AI rather than a document thrown over a wall. Design stays human, because trade-offs (consistency versus availability, build versus buy) depend on business context. Tests and evals become the contract that tells the AI what "correct" means. And legacy code that was once "too risky to touch" becomes safe to modernize.
Who does what now: conductors, orchestrators, and the harness
Developers move between two modes. In conductor mode they work in real time with the AI, reviewing each change as it appears, which suits complex or unfamiliar code. In orchestrator mode they define a task, hand it to one or more agents running in the background, and review the result later, which suits well-scoped work like bug fixes, migrations, and test generation. The same person uses both in a day.
The model itself is only one part. Around it sits the harness: the rule files, tools, sandboxes, and checks that let the model finish real work. A useful equation from the paper is Agent equals Model plus Harness. This matters for a plain reason. When an agent gets something wrong, the instinct is to blame the model, but most failures trace back to a missing tool, a vague rule, or a missing check. Getting the context right (the specs, conventions, and constraints you feed the agent) does more for quality than swapping models. Treat those context files as team assets: reviewed, versioned, and owned by named people.
What it costs: the CapEx and OpEx flip
For a project manager, the money story is the one that lands in a budget. Vibe coding looks cheap because the upfront cost is near zero. The hidden cost is operational. Every prompt spends tokens, and a loop of "fix your own mistake" burns tokens with a low first-pass hit rate. Unstructured AI code becomes a maintenance tax when a bug surfaces months later, and unverified code becomes a security bill in production.
Agentic engineering reverses this. You pay more upfront to design specs, tests, and context, and the marginal cost of each new feature drops. Two levers keep running costs down. Context engineering means sending a dense, relevant payload instead of the whole codebase every time. Model routing means sending hard tasks to a large model and cheap, routine tasks (a unit test, a small review) to a smaller, cheaper one. The trade is a familiar one: higher capital cost now for lower operating cost later.
Three things for project managers to hold onto
- The 80% problem is an estimation trap. An agent can produce roughly the first 80% of a feature quickly. The last 20% (edge cases, error handling, integration, correctness) needs judgment the model often lacks, and that is where the real time goes. Plan schedules around the hard 20%, not the easy 80%.
- Velocity is not the same as speed. In METR's controlled trial, experienced developers using AI took 19% longer while believing they were about 20% faster (METR). Time shifts from writing to reviewing and correcting. Measure delivered, verified work, not lines produced or how fast it feels.
- AI amplifies your engineering culture. Teams with strong tests, clear standards, and real code review get much more from AI. Teams without them ship faster into a bigger mess. Adoption alone is not a strategy: usage hit record levels in the Stack Overflow 2025 survey even as trust in AI output fell.
FAQ
Is vibe coding safe for production?
Not on its own. It is ideal for prototypes and internal tools. For anything customers depend on, add the specs, tests, and review that define agentic engineering.
Will AI make my team smaller?
It changes the mix more than the size. Smaller teams can take on larger problems, but the scarce skill shifts to specification, evaluation, and architecture, not typing code.
What is the single best first step?
Set up a project rules file (an AGENTS.md or equivalent) with your stack, conventions, and hard rules, then write tests and evals before generating code. Together they are the contract with the AI.
How do I explain the risk to leadership?
Say the team practices agentic engineering, with AI writing code under human-designed constraints and test coverage confirming correctness. That is a very different conversation than "we are vibe coding the payment system."
Sources
- Andrej Karpathy, original "vibe coding" post (Feb 2025): https://x.com/karpathy/status/1886192184808149383
- Vibe coding, overview and reception (Wikipedia): https://en.wikipedia.org/wiki/Vibe_coding
- METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (Jul 2025): https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- Stack Overflow 2025 Developer Survey (AI adoption and sentiment): https://survey.stackoverflow.co/2025/
- Grounding material: Osmani, Saboo, and Kartakis, "The New SDLC With Vibe Coding," Google (May 2026)
Top comments (0)