Learning how to BMAD

A developer documented a spec-driven, multi-agent workflow called BMAD for scaling agentic software development, in which a colleague named Vlad used the framework to refactor a monorepo CI/CD pipeline that had 90-minute test cycles, completing the work asynchronously in a single weekend. The account recommends capping working context between 100k and 260k tokens because larger windows degrade output, avoiding git submodules in favor of umbrella monorepos, and building a custom BMAD skill to mechanically validate traceability across PRDs and ADRs. The developer argues the bottleneck in AI-driven development has shifted from implementation speed to cognitive load and context window management.

My evolution in agentic development has been somewhat typical. At the end of 2025, it was all about the new thing at the time — Vibe Coding. Then, at the beginning of 2026, SDD with OpenSpec became a big game changer. It provided a way to scale agent-driven development without letting the codebase turn into a mess. And the next natural step? Multi-agent agentic workflows with BMAD. At first glance, BMAD seems vast, sophisticated, and complex. So where do I start? The best approach I could come up with was pretty simple: I talked to a friend who has been using BMAD for a couple of months, transcribed our conversation, and asked GenAI to summarize it. Here’s what I got. Software development constraints have shifted. The bottleneck is no longer keystrokes or raw implementation speed; it is cognitive load and context window management. My colleague Vlad recently scaled an entire production architecture using the BMAD AI framework. By shifting his role from writing syntax to orchestrating architecture, he maxed out Claude’s highest subscription tiers and fundamentally changed his output vector. Tasks that typically rot in a backlog are now executed in a single weekend. Here is the operational framework for high-leverage execution using BMAD. AI-driven development requires strict adherence to spec-driven workflows. You do not ask the AI to "build a feature." You force it to define the product, register the architecture, and validate its own readiness before a single line of code is written. The Spec-Driven Workflow: Brainstorming → PRD → ADR → Epics → Stories → Readiness Validation Loop if Not Ready → Implementation → Code Review Under legacy models, technical debt persists because the ROI of fixing it rarely justifies the engineering hours. AI reverses this calculus. Vlad’s monorepo CI/CD pipeline was choking on batch scripts, pushing test cycles to 90 minutes. Rather than manually rewriting it, he orchestrated an automated intervention: A refactor that would have cost a week of human engineering time was resolved asynchronously while the engineer managed top-level logistics. If you are scaling an AI agent like BMAD, adopt these technical guardrails: Avoid git submodules and multi-repository architectures. Umbrella projects with submodules introduce heavy friction when tracking commit hashes across dependencies. AI agents perform best when they have unified line-of-sight across the entire codebase—front-end and back-end alike. Larger context windows yield higher hallucination rates. While Claude and local models like Qwen boast 500k to 1M token windows, pushing past 400k guarantees degradation. The AI will fabricate interruptions or lose the thread entirely. Cap your working sessions between 100k and 260k tokens. Isolate execution by breaking stories down further rather than bloating the context. As your project scales to hundreds of PRDs and ADRs, manual validation becomes impossible. Vlad engineered a custom BMAD skill using an open-source requirement flow tool. The skill mechanically sweeps all files to ensure strict traceability—flagging orphan PRDs, missing dependencies, and architectural contradictions before the implementation phase begins. Address structural unknowns and technical spikes in the earliest stories of an Epic. By resolving ambiguities at the start of the cycle, the tail end of the Epic becomes highly mechanical, requiring near-zero human intervention. We are exiting the era of the 50-person engineering team split across 10 specialized layers. The future belongs to single-engineer monorepos, where the human acts as the Chief Architect and the AI executes the labor. Focus on the architecture; the speed will follow.