Why Maestro moved away from BMAD: our decision and test results

We replaced BMAD with our own method to organise Maestro’s steps more directly. A campaign of nine scenarios in August 2026 informed that decision without establishing a general ranking of tools.

Maestro started with BMAD-METHOD, an open method for organising AI agents’ work. We wanted more direct control over the sequence: what the agent should produce, what the user approves and what happens when a check fails. We therefore developed our own method and then evaluated its results.

The test that informed our decision

On 6 August 2026, we evaluated our method v2.5 on nine scenarios previously used in our tests: inventory, shared expenses, room bookings and others. The run went from the initial need to a first development batch, with checking and repair. It did not aim to finish the entire application.

For comparison, we used nine BMAD-METHOD v6.10 outputs retained from earlier campaigns on the same requirements. Each anonymised output received three assessments by AI judges using shared rubrics. Our team knew the scenarios; these were not nine new exercises held back for an independent evaluation.

The scores, including the less favourable result

Our method received a higher score on eight scenarios. BMAD retained the lead on room bookings: 90.7 versus 89.3. Its design documents were sometimes more complete. That difference matters, but the scores combined design and implementation: they do not say how many features were ready for use.

What we changed in Maestro

Choosing an in-house method let us work on clearer steps and observed defects: a written rule missing from the code, an overlooked duplicate case or a check that did not launch the real workflow. The aim was to make corrections easier to locate within the agent team’s work.

That does not make model responses predictable or software error-free. A method defines an organisation; the quality of the result still depends on the stated need, the implementation and the checks actually performed.

The limits of this comparison

We designed this protocol and improved our method using the same scenarios. The BMAD outputs came from earlier tests, with their own conditions and spending caps. One retained run per scenario and method does not measure consistency. Small scoring differences should therefore remain observations from this campaign.

Our subsequent tests also revealed highly rated software that did not start correctly. We describe these defects in the campaign review. These historical results do not prove that Maestro outperforms BMAD today, or other application-building tools.

What this means when choosing a tool

Look at how a tool helps you check your project: find a decision, try a feature, correct a defect and inspect what changed. Our measurement of instruction volume describes another aspect of our method. It is no substitute for trying the tool on your own needs either.