SDLC and ADLC: The Two Vibe Coding Processes You Need to Ship Real Products That Run Safely Long-Term

Quick answer

• SDLC (Software Development Life Cycle) is the traditional software development process with six steps: Planning/Requirement, Design, Code, Testing, Deploy and Maintain. According to the original author, skipping any of these six steps leaves a vibe-coded product at demo level, not ready for a real environment.
• ADLC (Agent Development Life Cycle) is the agent development life cycle, in which an AI agent performs most of the process itself: it plans, designs, writes code, evaluates itself against a test set, fixes failures, deploys and monitors. According to the author, this is the mechanism that lets tools such as Claude Code, Codex and Cursor run overnight.
• The hybrid model the author uses combines both processes: humans set out the idea and approve the spec (SPEC), while the agent writes code, self-tests, deploys to a staging environment and to production via CI/CD, then monitors and learns from errors. The effective loop is Human ↔ Agent, not an agent running 100% on its own.
• According to the author, most vibe coding projects fail because they lack a process, not because of poor prompts. The figures and opinion-based claims in the original post are the author's personal experience shared with a community, not independently verified data.

What is SDLC?

SDLC (Software Development Life Cycle) is the standard process for building software from idea to operation, with connected stages: planning and requirements, design, coding, testing, deployment and maintenance. Each stage produces a concrete output that becomes the input of the next, so product quality can be controlled before the software reaches real users.

What is ADLC?

ADLC (Agent Development Life Cycle) is a software development process whose main executor is an AI agent. As the original author explains it, ADLC takes the SDLC frame and replaces "S" (Software) with "A" (Agent): requirements go in, the agent plans, designs and writes the code itself, evaluates the result against tests supplied by humans (eval), loops back through code and test when it fails, deploys itself when it passes, monitors itself in operation, and returns to the design step when a new feature appears.

What are the six SDLC steps in vibe coding practice?

According to the author, whether you code by hand or vibe code, a real project still has to pass all six steps:

1. Planning / Requirement: spell out the problem and the business flow, usually as a business flow file. This is the input that lets the AI understand what you actually want instead of guessing from a few scattered prompts.
2. Design: draw the interface and prepare dummy data so you can see what the finished project will look like.
3. Code: choose the tech stack, set up supporting skills and add instruction files such as CLAUDE.md so the agent works to the project's conventions.
4. Testing: test the whole codebase, including security and performance. According to the author, without this step the product is not ready for production.
5. Deploy: ship the product to a real environment, usually a VPS for a real project. The author argues serverless should only be used for small landing pages or personal projects — this is the author's personal opinion, not a mandatory rule for every project.
6. Maintain: monitor the running application, observe real behaviour and fix issues that appear.

https://pomf2.lain.la/f/so3dvzn4.webp

How does the ADLC eval loop work?

Unlike SDLC, which proceeds version by version, ADLC is a non-stop loop, as the author describes it:

• The human feeds the requirements into the system.
• The agent plans, designs and writes the code itself.
• The agent evaluates itself against the test set the human provided (eval): on failure it loops back and rewrites the code; on success it moves on to deployment.
• The agent deploys to a versioned environment, monitors itself in operation and learns from errors.
• When a new feature arrives, the loop returns to the design-spec step instead of starting over.

According to the author, the only drawback of pure ADLC is token cost, because every code–test–fix loop consumes model resources.

https://pomf2.lain.la/f/fv1kbx25.webp

How do SDLC and ADLC differ?

Criterion	SDLC	ADLC
Main executor	Humans (the development team)	AI agent, with humans steering
How phases proceed	Linear, version by version; each phase must finish and be approved before the next	A continuous loop; quality is measured by eval after every round
How errors surface	Errors found late are expensive to fix because earlier phases must be revisited	Errors are caught early by automated eval; the agent fixes and reruns itself
Human role	Does and approves each step directly	Shapes goals, writes and approves the spec, accepts the result
Distinctive cost	Time and staffing per version	Tokens for the agent's loops (according to the author)
Traceable output	Documents and artifacts from each phase	Artifacts from each step plus eval results from each loop

How does the hybrid SDLC + ADLC model split work between humans and AI?

The model the author uses keeps direction and acceptance gates on the human side, and execution and iteration on the agent side:

Step	Who	What
1	Human	Set out the idea
2	AI	Generate the SPEC file from the business flow file (using a skill)
3	Human	Approve the SPEC — a mandatory step, never skipped
4	AI	Write code with all the CLAUDE.md and AGENTS.md files needed
5	AI	Re-read the requirements and run the real-scenario test set
6	AI	On failure, loop back through Code + Test
7	AI	On success, push to the VPS for human acceptance
8	Human	Accept the build on the VPS
9	AI	Push to production via CI/CD
10	AI	Monitor, learn from user-reported errors, return to SPEC for new features

https://pomf2.lain.la/f/61zhmpq2.webp

The core of the hybrid model is that every step leaves an artifact — the business file, the SPEC, the code, test results, the deployment — so the whole process can be traced when something goes wrong. The author sums it up with the line: "Prompting is for discovery, Declaring is for delivery" — prompting fits the exploration phase, while delivering a real product needs clear declarations such as a SPEC and a test set.

Why does understanding the eval loop make overnight agent runs safe?

According to the author, what lets agent tools run for long stretches without "coding nonsense" is not longer prompts but the eval loop: the agent may only proceed when the human-defined test set confirms the result passes. Without a trustworthy test set and a human SPEC-approval gate, letting an agent run 100% on its own still risks producing code that looks right but is wrong for the business. The author also notes that a sample document set — business flow file, CLAUDE.md and the VPS acceptance plus CI/CD process — is linked in the first comment of the original post for readers to reference.

https://pomf2.lain.la/f/ymvtps3f.webp

Frequently asked questions about SDLC and ADLC

Can vibe coding skip the Testing step of SDLC?

It should not. According to the original author, Testing — testing the whole codebase, including security and performance — is the condition for a product to reach production; without it, the product stays a demo. In the hybrid model the agent runs these tests itself against the human-provided test set, but the human still performs final acceptance.

Does ADLC completely replace SDLC?

Not in the way the author presents it: ADLC uses the SDLC frame as its foundation and swaps the executor for an agent. The core stages — requirements, design, testing, deployment, maintenance — still exist; what changes is that the agent repeats them faster and continuously.

Why does pure ADLC cost so many tokens?

Because the agent runs every loop itself: each failed eval means re-reading context, rewriting code and rerunning tests, and all of those loops consume tokens. That is why the author chose the hybrid model, where human SPEC approval reduces unnecessary loops.

Which model should a beginner adopt?

According to the author, the hybrid model is the balanced choice: start from the idea, let the AI generate the SPEC, review the SPEC carefully yourself, then let the agent code and self-test, and finally accept the build on the VPS before the agent pushes to production via CI/CD. Beginners do not need to let the agent run everything from day one.

Can serverless be used for real projects?

The original author argues serverless should only be used for small landing pages and personal projects, while real projects should be deployed on a VPS. This is the author's personal view and experience; in practice, infrastructure choices also depend on each project's scale, budget and operating requirements.

Sources

This article was rewritten from a post by Code Web Không Khó on Facebook (published 4 November 2026): the original Facebook post. The original post comes with an illustrative reel (facebook.com/reel/1418764769973870) and points to a document set in its first comment/pinned post. Claims such as "most vibe coding projects fail for lack of process", ADLC being the process large companies use for agents, and the serverless recommendation are the author's personal views and experience in a community post, not independently verified data. The illustrations are the SDLC vs ADLC infographic slides taken from the original post itself.