We now develop only with AI. How our Smithy works

At bmpi, new software is now built only with AI. The agents write the code, the tests and the documentation. Our developers describe what is to be built, decide on the approach and check the result. We work with Claude Code by Anthropic.

This shifts the work. Code is written quickly. What matters is that it is clear beforehand what is to be built, and that what was built is checked afterwards. For this we developed a process of our own: the Smithy.

Four phases

The Smithy has four phases. Two of them frame the whole project and run once each: Ignite at the start, Seal at the end. In between, every single task goes through Forge and Temper. Each step is a command of its own with a fixed result that the next step takes over.

Ignite: lighting the fire

Once per project, we lay the ground.

• Draft reads the existing prototype and draws up the backlog with epics, tasks, dependencies and acceptance criteria. First comes a walking skeleton: one thin thread through every layer, from the interface to the database. Only when it stands does the work broaden.
• Blend decides which qualities the application must have. The basis is the ISO/IEC 25010 standard; the stakeholders weight the characteristics themselves.
• Raise sets up the repository, the build pipeline, the templates for architecture decisions and the working instructions for the agents. This includes a quality gate that checks every change.

Forge: forging each task

Every task goes through three steps, with checkpoints in between.

• Spec turns a rough idea into a ticket with clear requirements: what is built, and how you know it is done.
• Design works out the solution and breaks it down into executable steps.
• Build turns the plan into code, step by step, with a log for each step.

A person reads the ticket and the plan before the first line of code is written. That is where the important decisions are made.

Temper: hardening and testing

At the end of each task, the result is hardened, tested and recorded.

• Assay reviews the change thoroughly: with the same checks as the pipeline (architecture against the documented rules, security, mutation testing), and also for business correctness, gaps in the tests and performance. Confirmed findings become new tasks for Build. This repeats until the review is clean. After three rounds, a person decides.
• Inscribe records what the task taught, in the shared memory (see below).
• Carve writes down the architecture decision: what was decided, and why.

Seal: closing the project

Once, at the end, we take stock. Reckon produces the retrospective from the logs of all tasks. Refine consolidates the architecture decisions into one consistent version. Chronicle writes the project documentation.

The memory: Engram

An agent starts every session with no memory of the last one. That is why there is the Engram, a shared knowledge store for all projects. It has two doors:

• In, on purpose. At the end of each task, Inscribe records the insights, selected and worded with care.
• Out, by reflex. Before any larger piece of work, the agent searches the Engram for what has already been learned about it.

So every session is smarter than the one before. What a team finds out in one project is available to the next.

What this means for our clients

• Clarity before the first line of code. Every task has requirements and acceptance criteria that a person has read.
• The same checks for every change. Architecture, security and tests run for every task, locally and in the pipeline.
• Traceable decisions. Every architecture decision is documented with its reasons.
• Reliable figures. Our standards say: calculations stay deterministic. In an application, an AI agent may query and present results; it may not change them. For reporting and the investment process, this is the foundation.

Our specialists are responsible for the result. The Smithy gives them more time for what counts: understanding the problem and checking the solution.

We are glad to show you how the Smithy can work in your project: Consulting at bmpi.