Two approaches to the agentic SDLC
Two approaches to the agentic SDLC are starting to emerge. They rely on the same coding agents and yet place the center of gravity in two different places.
A workflow left free
In the published practices around the SDLC with Claude Code, Anthropic deliberately leaves a lot of freedom to the developer. The workflow is not imposed and the project documentation remains relatively free in its structure.
This flexibility relies largely on the model's capabilities: exploring a codebase, understanding the context, drawing up a plan, producing the code, testing, then iterating with the developer. The quality of the result then depends heavily on the model's ability to rebuild the context, and on the developer's ability to guide it.
An SDD embedded in an SDLC
An SDD approach embedded in an SDLC, such as BMAD and its Niji ARC™ variant, shifts the center of gravity much more clearly. The specification becomes one of the main supports of delivery.
| Dimension | Free workflow | SDD embedded in the SDLC |
|---|---|---|
| Center of gravity | The model and the developer | The specification |
| Workflow | Not imposed | Organized through workflows and skills |
| Project documentation | Relatively free structure | Shared reference: business needs, journeys, architecture, decisions, acceptance criteria, tests, RUN |
| Who contributes | Mainly the developer, with the model | Business, Product Builder, project management, designer, architect, developers, QA |
| Where quality rests | On the ability of the model and the user to rebuild the context | On a shared framework |
| Strength | Quickly benefiting from model progress, strong developer autonomy | Delivery consistency and continuity |
The specification, a shared reference for the project
In this second approach, the specification becomes a shared reference for the project, beyond the mere context intended for agents. It brings together:
- business needs;
- user journeys;
- architecture;
- technical decisions;
- acceptance criteria;
- tests;
- RUN constraints.
Business, Product Builder, project management, designer, architect, developers and QA can thus work from the same material, which agents can also use. Everyone contributes from their own role, and what one produces directly feeds the work of the others, humans and agents alike.
This is the principle we apply with Niji ARC™: the specification is the product, and agents generate the code, tests and documentation directly from it. The Product Builder's role in this system is detailed in our thinking on AI squad size.
Workflows and skills: organizing the collective build
Shared material is not enough: the way it gets built also has to be organized. That is the role of workflows and skills, which define who contributes, at what moment, with what level of context and which controls.
Each role steps in from its own scope: business, product, design, architecture, development, QA. Workflows make these contributions explicit instead of leaving them to each person's initiative.
Contributions are sequenced along the cycle, from the expression of the need to going live, with steps identified end to end.
Each agent and each contributor receives the context useful for their step, drawn from the shared reference, without having to rebuild it.
Validation points mark out the path: acceptance criteria, tests, reviews. The material produced at one step becomes the basis for the next only once it has been checked.
We apply the same logic of roles, skills and workflows to data professions with the MAD framework.
The difference shows at industrial scale
The difference becomes particularly visible when you try to industrialize.
A very free workflow lets you quickly take advantage of model progress and keeps strong autonomy for the developer. For a developer or a small team, that is a real asset.
A more structured SDLC instead seeks to rest a larger share of quality, consistency and delivery continuity on a shared framework, rather than on the sole ability of the model and the user to rebuild the context.
Three questions to place your context. How many people and roles contribute to the same product? What is the product's lifespan, and who will keep it running in RUN? When a contributor moves to another project, does the context stay in the reference or in someone's head?
From agentic coding to agentic delivery
The stakes are then no longer only the efficiency of agentic coding. They become those of agentic software delivery, in which humans and agents contribute to the same reference system, each from their own role.
Models will keep improving. The question of which reference humans and agents build on together will arise whichever model is used.
Key takeaway. A free workflow takes advantage of the model's power. A structured SDLC makes the specification a shared reference where humans and agents contribute, each from their own role.
