Photo by Luca Bravo on Unsplash
Most devs using AI coding tools know the drill. You write a vague prompt, get some code back, tweak it until it runs, and hope for the best. It’s fast. It’s exciting. But later, someone has to untangle that rushed mess—or worse, they don’t. Enter the dreaded “vibe coding” hangover.
But Codev, a new open-source framework, wants to change that story.
From messy prototypes to solid software
Let’s break it down: Codev was built for developers who want the speed of AI without the chaos. Instead of treating your conversation with an AI as a throwaway step, Codev makes that discussion part of the actual software — like, literally part of the codebase.
The key idea? Words matter. That high-level “let’s build a todo app” chat you have with the AI? Codev captures it, turns it into structured, reviewable steps, and makes it version-controlled.
Think of it this way: instead of writing code, you write specs and plans in natural English—and the AIs (yes, plural) take it from there.
What is Codev really doing?
Photo by Desola Lanre-Ologun on Unsplash
At the heart is something called the SP(IDE)R protocol. It’s a step-by-step process where human devs and multiple AI agents work together to move from idea to finished software.
Here’s how it flows:
- Specify: You and the AI define what you’re building and how you’ll know it works.
- Plan: AI maps out a phased implementation.
- IDE loop: For each phase, an agent:
- Implements the feature
- Defends it by writing tests
- Evaluates if it meets the spec
- Review: Final pass where lessons are documented — and future agents can get smarter.
It’s collaborative, detailed, and downright disciplined.
And those different AI agents? Each has its strengths.
Waleed Kadous, Codev’s co-founder, pointed out that Gemini’s great at spotting security flaws. In one test, it caught a cross-site scripting vulnerability and flagged a bug that would’ve leaked an API key — a mistake that could’ve cost real money. Meanwhile, GPT-5 shines at simplifying tricky designs.
But no workflow is fully automated. Every step still includes checks by a human engineer. It’s AI-augmented, not AI-run.
Yes, they used Codev to build Codev
This might be my favorite part: the team actually “dogfooded” their own system. They used Codev to build itself.
That including setting it up. There’s no complex install wizard—just ask your AI to apply the Codev GitHub repo to spin things up. The AI decides how to set everything up intelligently based on your project.
As Kadous puts it, “Natural language is executable now. The agent is the interpreter.”
Does it actually work?
To see Codev in action, the team ran a test:
They gave Claude Opus 4.1 (yep, an AI model) the same prompt twice: build a modern web-based todo app.
- Run 1, using standard “vibe coding,” produced a nice-looking demo—but under the hood, it was bad. No database, no API, no tests, and it didn’t meet any of the requirements.
- Run 2, using Codev’s process, generated:
- 32 source files
- A SQLite database
- Full RESTful API
- Five test suites
- 100% functional match to the spec
Better yet, the human developers didn’t touch a single line of code themselves.
Subjectively, Kadous says he feels 3x more productive using Codev. And when he ran the output past other LLMs to judge quality, they said it looked like something a well-oiled dev team would’ve built.
So what changes for developers?
Photo by Shamin Haky on Unsplash
It’s not about writing lines of code anymore. It’s about writing smart specs and structured plans—then reviewing what the agents come up with.
For seasoned engineers, that means focusing on higher-level thinking, applying background knowledge, catching edge cases, and shaping direction. A spec or planning phase could take an hour or two, but it’s worth it when the AI handles the elbows-deep coding work.
Kadous says, “The people who will do the best… are senior engineers and above because they know the pitfalls.”
But it doesn’t mean junior devs are left out. If anything, we might need to start thinking about how we train and mentor them in this new setup — giving them the chances to build those architectural instincts early.
A smarter kind of collaboration
Codev paints a picture of what the next wave of AI-assisted software development could look like: collaborative, structured, human-led, and AI-enhanced.
The chat you have with AI doesn’t vanish into the ether. It’s part of your codebase. Auditable. Maintainable. Versioned. And the code it generates isn’t throwaway—it’s ready for production.
In this world, your words matter just as much as your code.
Keywords: Codev, vibe coding, AI coding tools, SP(IDE)R protocol, AI software development, AI agents, natural language programming, open-source frameworks for developers, GPT-5 code generation, Gemini security review, enterprise coding tools, code documentation
