Why we left BMAD: an autopsy in numbers

Maestro started with BMAD-METHOD. We left it for our own engine, Partition, and measured the replacement campaign after campaign: an average of 89.5 versus 83.5 across nine tasks, with eight duels won out of nine.

Maestro started with BMAD-METHOD, the most installed open agentic framework in the field. We left it for Partition, our deterministic playbook engine, and measured the replacement instead of simply asserting it: in a campaign of nine realistic tasks judged in parallel, an average of 89.5 against BMAD's 83.5, with eight duels won out of nine.

What BMAD does well

It must be said honestly: BMAD's design is rich. Its PRD is often more extensive than ours, its stories more detailed, its ceremonies intended to cover broad ground. On one task in our campaign, a room-booking API, it even won a duel, 90.7 to 89.3. It is not a weak opponent, and the protocol always treated it fairly: the same rubrics, the same judges, never a task prepared in advance on either side.

Where it breaks

The gap widens in execution, not design. Across the nine tasks in one campaign, BMAD exhausted its budget three times without crossing the finish line: it spends on documents and dies before code. When code exists, it remains thin: on a shared-expenses project for housemates, only one of six stories was actually coded. Tests, when present, are rarely truly run. Criticisms of BMAD in 2026 converge on the same point: a prescriptive workflow whose formality outweighs its gains, fixed personas that weigh more than they help. Those personas are precisely a design choice that makes sense on paper, one agent plays the architect, another the writer, a third the tester, but in practice adds a layer of negotiation between agents without ever guaranteeing that the final result holds up.

What the migration cost, what it returned

We did not merely switch engines: we measured every Partition version against BMAD, with the same rubrics and judges and new tasks in every campaign. In a four-task campaign, Partition delivered at 47% of BMAD's cost, with BMAD averaging 84.2. In a later eight-task campaign, the gap widened further: 93.3 versus 85.6, seven duels won out of eight. The full details, including the acknowledged initial defeat, are in our measured duel over six days.

What a deterministic playbook engine changes

The fundamental difference is structural, not philosophical. BMAD leaves substantial room for improvisation by multiple agents negotiating with each other. Partition sequences compact playbooks, scoping, specification, architecture, breakdown, development, verification, each with a precise output contract, each approved by you before the next. Tests are written before code, bounded repair corrects failures, and nothing is repaired in an endless loop: beyond two attempts, the decision goes back to the human instead of going in circles.

The real lesson

Every Partition version that beat its predecessor arose from a defect found by the previous version's jury, never a design intuition. It is the same discipline we apply when an agent plays our most difficult user: measure before asserting, change one rule at a time, exactly where it failed. Leaving BMAD was not a judgement on its authors; it was the result of repeated measurement.