Skip to content
Technology transformation expertise firm · ARDIANE Group Innovate, secure, succeed

Artificial intelligence

Should you try the BMAD method to organise AI-assisted development?

In brief

The BMAD method is a third-party open project, published on GitHub under the MIT licence, that organises AI-assisted development around role-specialised agents producing documents that the next ones take over. Management Square presents its content, what it requires, the cases in which to try it and a way of judging a trial, based on the project's repository and documentation.

Page contents

A code assistant receives a request and returns code: nothing in this exchange requires the decisions taken along the way to be written down. The BMAD method starts from this observation. It is not a Management Square method: it is a third-party open project, published on GitHub under the name BMAD-METHOD, that organises AI-assisted development around role-specialised agents (analyst, product manager, product owner, architect, developer, designer), each of which produces documents that the next ones take over.

Trying it can be justified when several people must agree on what a product should do and decisions must survive from one session to the next. An isolated fix does without it. We describe the content of the framework, what it requires, the cases in which to try it and a way of judging the trial, based on the project’s repository and documentation read on 12 November 2026. Nothing is asserted about its results.

Role-based agents, documents handed on and a build loop

The repository belongs to the GitHub organisation bmad-code-org. The MIT licence, in the name of BMad Code, LLC, covers the software; the names BMad and BMAD-METHOD are trademarks of the same company, excluded from this licence according to the repository’s TRADEMARK file. The README does not spell out the acronym. It presents the project as “Agile Ai Driven Development”.

At the date of reading, two lines coexisted: the stable version 6.12, installed by an npm program and frozen except for fixes, and a version 7 in alpha preview, whose README announced its release for mid-November 2026.

The framework is installed in the form of “skills”, named commands added to an assisted programming tool. The version 7 documentation describes six named agents, loaded by these commands: Analyst, Product Manager, Product Owner, Architect, Developer and UX Designer. The 6.12 branch has five, without the Product Owner.

The documentation arranges the preparation commands “from analysis through planning to solutioning”: exploring an idea, research, a scoping brief, a product requirements document (PRD), user experience, architecture. Then comes the build: the bmad-build command examines the code and the upstream documents, plans the change, writes it, submits it to independent reviewers and corrects what they point out. Finally, a retrospective judges a set of stories (epic) against its acceptance criteria.

These agents are not always distinct programs. In the default multi-agent discussion mode, a single model lends its voice to all the personas; the documentation observes that it then tends to make them agree, and offers modes that launch separate agents, at higher cost. The nuance matters for anyone who would read the framework as a multi-agent system in the strict sense.

Written decisions, in exchange for an installation and discipline

The project claims a process proportionate to the work and a lasting context: product and technical decisions are written once, then reused instead of being re-explained in each conversation. What can be checked during a trial is simpler: each step leaves a document in the project, code reviews are planned, and team settings are stored in a folder versioned with the repository, so that everyone inherits the same conventions.

In return, the framework requires several things.

  • A workstation on which Node.js, Python and the uv tool can be installed, and a programming tool that accepts skills.
  • People available to answer the agents’ questions and review their documents. The documentation warns that a wrong answer to an open question is the most costly error to discover later.
  • A person who arbitrates the product, a role close to that of the product owner: according to the documentation, when the work belongs to an organisation, the PRD is the document it owns, and nothing downstream reinterprets it.
  • A review of permissions. The installation page indicates that the skills run Python scripts, that customisation files can add steps, that some read texts written by others (tickets, merge requests) and that they act with the tool’s permissions. These points fall under the controls described in our article on AI agent security. A customisation can also point an agent towards MCP servers: each one widens what it can reach.
  • A tolerance for change, if the preview version is chosen: its README announces breaking changes between updates.

When a trial is justified, and when a lighter approach is enough

The documentation itself draws a line: an obvious, low-risk change does not need the framework. It reserves the full path for cases in which people who did not carry out the thinking must approve the product, in which several sets of stories, teams or agents build on the same decisions, and in which a regulator or a steering committee requires named documents.

We suggest adding two conditions before a trial: an application the team knows, to distinguish what comes from the framework from what comes from the subject, and the agreement of security on running scripts from the project. A trial covering an entire application remains a software project, whose running our custom software development page describes, from scoping to takeover by the teams.

Two simpler routes exist. The approach described in our article on spec driven development chains specification, plan and tasks without dividing the work among roles; it may be enough when the need is already clear and carried by a single person. Using an assistant without a framework suits changes that a quick review is enough to check, under the rules proposed in our article on AI-assisted development. We suggest basing the choice on the number of people to bring into agreement and on the lifespan of decisions, rather than on the size of the code.

One confusion should be ruled out. BMAD’s agents help build software; they are not the AI agents that carry out actions in the enterprise’s processes, whose authorisation is covered in our article on enterprise AI agent use cases. The design of the latter, tools, permissions and supervision included, falls under our AI agents page.

Judging a trial on measures set before installation

We suggest limiting the trial to one team, a limited set of stories and the stable version, with an end date fixed at the outset. Baseline measurements are taken on this team before installation. The measures our article on AI-assisted development proposes for an assistant apply here; the framework calls for others.

On the input side: the number and length of documents produced before the code, the time spent answering the agents and reviewing, the number of build sessions, the cost of model usage. On the quality side: the discrepancies found by human review after the agents’ review, the defects discovered in acceptance testing, the changes abandoned along the way, and a simple durability test, which consists of asking an outside person to find in the documents the reason for an architecture choice.

The criteria that would lead to extending, restricting or stopping are written down before the start. Restricting can mean keeping only the preparation commands: the documentation notes that some users employ only these, without ever entrusting the code to the framework.

FAQ

Frequently asked questions

Another question? Ask us

What does the acronym BMAD stand for?

The README of the BMAD-METHOD repository, read on 12 November 2026, does not spell out the acronym. The file that describes the project's trademarks lists among them the slogan "Build More, Architect Dreams", whose initials form the name, without presenting it as its expansion. The prudent course is to use the name as it stands.

Is the BMAD method free and open to use?

The software is published under the MIT licence, in the name of the company BMad Code, LLC, and the README describes it as free and open. The names BMad and BMAD-METHOD, however, are trademarks of that company, which the MIT licence does not cover according to the dedicated file in the repository. These conditions should be reread on the date of adoption, as they may change.

What is the difference between the BMAD method and spec driven development?

Spec driven development has a specification, a plan and tasks reviewed before the code assistant writes. The BMAD method is a specific open framework that divides this work, and the documents that precede it, among agents each holding a role. The second can contain the first without being reduced to it.

A project to bring to completion?

A first conversation is enough to understand your challenge, identify the right people and tell you frankly what we can bring.