Two obstacles make "give the model everything" fail: long-context degradation and the absence of real memory over time.
Long-context degradation
The LoCoDiff benchmark (January 2025) shows that even the best available model, Sonnet 4.5, degrades sharply with long contexts.
What happens: accuracy drops from 96% with 2K–21K-token contexts to 64% beyond 60K tokens. When you hand the whole codebase to the model, it drowns in noise.
•State-of-the-art models still rely on manageable context windows to keep high accuracy.
•Feeding entire files or repositories produces noise, repetition, and inconsistent decisions.
A METR study (July 2025) found that experienced developers working on their own projects were 20% slower using AI than working alone.
Why? The LLM restarts from scratch every time: it lacks the tacit context that humans accumulate. Experts predicted they would be 39% faster, but reality was the opposite.
•Observed time per story: 1.67 h without AI vs. 2.26 h with AI assistance.
•Expectations forecasted 24%–39% faster delivery, exposing the gap between predictions and reality.
The solution: an agentic team that plans and executes
BMAD is an AI development method where eight specialized agents work like a real team. They plan the whole project before coding and then execute story by story with continuous validation.
Each agent is the same LLM with a specialized prompt and verified, sharded context. No magic memory—just structured documentation within reach.
Sharding
Large documents are split into focused shards. Instead of handing the Developer a 10,000-token PRD, they receive a 400–600-token shard with only what matters for the story.
Result: every agent operates in the 96% accuracy zone (<5K tokens) without extra noise.
Specialized agents
Eight agents with concrete roles: Product Manager, Architect, Developer, QA, etc. Each has tailored instructions and checklists, and only queries the context they need.
Result: surgically precise context for every decision without overloading the model.
Structured documentation
Rigorous planning before coding: complete PRD, defined architecture, and sequential stories that validate each other. That creates the "memory" LLMs don't have.
Result: agents work with professional, verifiable information always aligned with the project goal.
Why BMAD works
1. Optimized context
In the METR study, teams worked with huge codebases in Cursor and ended up 20% slower. BMAD does the opposite: PM handles 2K tokens of requirements, Architecture ~3K tokens, and the Developer a single epic (1.5K tokens). High accuracy, no degradation.
2. Validation and transparency
The Product Owner validates every story draft before development, QA reviews the Developer's work, and humans approve each milestone. Artifacts stay human-readable and versioned in the repo.
3. Advanced elicitation
Agents ask questions and iterate on documentation. It's not just generating code—they discover, clarify, and update requirements while keeping the person in the loop.
Agentic team
The eight BMAD agents
Each role is the same model with a specialized prompt and sharded context. The outcome: a multidisciplinary team that works like a human crew.
Business Analyst
Insightful Analyst & Strategic Ideation Partner
/analyst *create-project-brief
Market research, brainstorming, competitive analysis, project briefs, and brownfield documentation.
