Vibe coding needs a pipeline, not a ban
Andrew Murphy

In short
- At some firms AI agents now write almost all the implementation code.
- Vibe coding is dangerous but banning it is throwing the baby out with the bathwater.
- Wrap engineering round it. Put the checks in the pipeline, with standard checks wherever you can and AI-based ones only where you need them.
- Give agents the jobs a team used to share and keep the gates out of their reach. The engineer moves to the what rather than the how.
This month AWS vice president Marc Brooker told GeekWire that for most Amazon developers "almost all of their implementation code is agent generated rather than human written". That is not a startup experiment. It is Amazon.
The latest Communications of the ACM carries an opinion column by Mayukh Mukhopadhyay asking whether "vibe coding" can grow up. His answer is VIBE: Verifiable, Interpretable, Balanced and Evolvable code, with tests on every round and regular refactoring. He leaves most of it to the developer.
I agree with the checks but I'm not sure I agree with asking the developer to do this. We've all noticed that "checking" Claude Code just becomes a monkey hitting either number 1 or number 2 repeatedly. Asking a dev to check line by line what an agent has written is unrealistic and unsustainable. It will work for the first five minutes and then it won't. Relying on the developer never worked for code people wrote by hand, so it really will not work better when the code arrives faster than anyone can read it.

Why a ban is the wrong answer
Some firms respond by banning AI coding tools, losing the gains along with the risks. Coding has always been the easy bit. The hard bit is the engineering round it and that is where engineers should be spending their time now. Automate the tits off it but automate it well.
We have been here before. Binary gave way to assembly, then high-level languages, then object-oriented ones such as Smalltalk and C++, then the frameworks we take for granted. Each step took developers further from the machine and each time we found new ways to keep software reliable, extendable and maintainable. Those practices started by hand and ended up automated. Vibe coding is the next step, only faster. This time the automation cannot wait.
Put the discipline in the pipeline
A pipeline does not get tired on a Friday and it does not care who or what wrote the change. Use standard, declarative checks wherever you can and save AI checks for where you need them. A standard check gives the same answer for the same code and costs a fraction of an AI call.
- Tests written from the requirement, not from the code. If the same model writes both, the tests may only prove it agrees with itself.
- Static analysis on every change. In a 2024 study nine models wrote 331,000 compilable C programs and formal verification found at least 62% of them vulnerable.
- Secret scanning. A key in the source is a key in the source, no matter who wrote the source.
- Dependency checks. A 2024 study found that on average at least 5.2% of the packages suggested by commercial models did not exist. For open-source models it was 21.7%. Anyone can register those names and wait for someone to install them.
- Gates the agents cannot touch. People control branch rules, build credentials and deploy permissions. An agent that can change its own checks has no checks. If a check can be skipped it will be, usually by someone senior in a hurry.
- A human sign-off and a practised rollback. The human agrees what the change should do and the tests that prove it. The pipeline checks the rest. When something gets through anyway, a practised rollback turns a bad change into an inconvenience rather than an incident.

Give the agents the team's jobs
A good team has a developer, a designer and a tester keeping each other honest. Agents can split the work the same way. One designs, another writes the code and a third writes the tests from the requirement. Nobody marks their own homework and the pipeline checks the result against the requirement. It is how we build our own website, ProbeSix and our internal tools. Agents on the same model can share a blind spot. A different model for the tests can help, though the real safeguard is that the tests answer to the requirement and not to the code. The engineer moves up to the what rather than the how.
Amazon is moving this way. Its revised guidelines emphasise measuring engineers by outcomes rather than inputs such as lines of code and it added mandatory peer review for production access after an AWS disruption it blamed on a misconfigured role. AWS chief executive Matt Garman said in April that Amazon is still "hiring just as many software developers as we ever had".
Software engineers have always been the smartest people in the room. Now they just have to do more engineering and less grunt work.
Monday morning
Check the protection on your main branch. Can a change reach production without passing its tests, a scan and a human sign-off? Can an agent change those checks or deploy straight to production? If either answer is yes, that was a problem before anyone said "vibe" and it is a bigger one now.
Simon Willison put it plainly: "Vibe coding your way to a production codebase is clearly a terrible idea." Put the vibes through the same gate as everything else.
Sources
GeekWire, Amazon rewrites its software engineering playbook for an era when AI writes the code (3 November 2026): https://www.geekwire.com/2026/amazon-rewrites-its-software-engineering-playbook-for-an-era-when-ai-writes-the-code/
Mayukh Mukhopadhyay, Can VIBE Drive Development? Threats and Opportunities, Communications of the ACM (November 2026): https://cacm.acm.org/opinion/can-vibe-drive-development-threats-and-opportunities/
Tihanyi et al., How secure is AI-generated Code: A Large-Scale Comparison of Large Language Models (2024): https://arxiv.org/abs/2404.18353
Spracklen et al., We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs (2024): https://arxiv.org/abs/2406.10279
Simon Willison's Weblog (6 March 2025): https://simonwillison.net/2025/Mar/6/vibe-coding/
