Implementation Tiers

Brain uses three implementation tiers. The tier choice for each feature is deliberate. It’s about matching the tool to the problem — using a full agent for a deterministic task wastes tokens and adds hallucination risk; using a skill for multi-step reasoning produces inconsistent results.

Tier 1 — Simple LLM + skill

A skill.md file with a clear prompt and small input/output contract. Stateless, deterministic-leaning, fast to iterate. Use for:

Generate a quotation with a fixed markup
Ask for missing client fields
Draft a quotation email
Summarize a single email
Classify an email topic

Pros: fast to build, easy to iterate, low engineering complexity, ideal for MVP. Cons: not suited for long-running tasks; limited state management; harder to coordinate multiple steps.

Tier 2 — LangGraph workflow

A multi-step state machine with explicit nodes for AI steps, tool calls, and human approval gates. Each transition is observable and resumable. Use for:

Email-to-style creation (intake → classify → extract → duplicate-check → approve → write)
Client creation approval flow
Duplicate check workflow
Quotation generation workflow
Any human-in-the-loop pipeline

Pros: clear states, good for human-in-the-loop, easy to debug, supports branching logic, strong observability. Cons: more engineering setup; requires workflow design up front; less flexible than full agent autonomy.

Tier 3 — Full agent with tools and environment

A reasoning loop with multiple tools, document access, image analysis, and the ability to plan multi-step actions. Used when the task can’t be reduced to a deterministic flow. Use for:

Tech pack → BOM (parse PDFs, parse Excel, vision analysis, garment template matching)
Manager lifecycle queries (multi-source retrieval, structured output)
Pricing intelligence (external search + internal history + client profile)
End-to-end oversight agent (continuous monitoring, risk classification, alerting)

Pros: handles complex analysis, multi-tool, multimodal; more autonomous. Cons: higher hallucination risk, requires stricter guardrails, needs strong observability, needs source attribution, more expensive to run.

Tier recommendation by feature

Feature	Tier
Fixed markup quotation	Tier 1 — Skill
Draft quotation email	Tier 1 — Skill
Missing field prompts	Tier 1 — Skill
Email topic classification	Tier 1 — Skill
Client creation	Tier 2 — Workflow
Manual style creation	Tier 2 — Workflow
Email-triggered style creation	Tier 2 — Workflow
Duplicate style check	Tier 2 + DB tool
True cost first draft	Tier 2 / Tier 3 hybrid
Tech pack → BOM	Tier 3 — Agent
Material matching	Tier 3 — Agent
Pricing intelligence	Tier 3 — Agent
Manager style lifecycle query	Tier 3 — Agent
Pipeline oversight	Tier 3 — Agent + scheduled monitoring
Approval inbox	Tier 2 workflow infrastructure

How tiers interact

A high-level workflow (Tier 2) often calls lower-level skills (Tier 1) as steps and delegates complex sub-problems to agents (Tier 3). For example:

Email-to-Style (Tier 2 workflow)
  ├── Classify email topic (Tier 1 skill)
  ├── Extract style fields (Tier 1 skill)
  ├── Duplicate check (Tier 2 with DB tool)
  ├── Show draft for approval (human gate)
  └── Write to CutMake (tool call)

After approval, next-step suggestion may chain to:

Tech Pack → BOM (Tier 3 agent)
  ├── PDF parse tool
  ├── Excel parse tool
  ├── Vision tool
  ├── Category template lookup
  └── BOM draft write

Each tier is the right tool for a different shape of problem.

Runtime — what hosts all three tiers
Workflows — see each feature’s tier choice in context

​Implementation Tiers

​Tier 1 — Simple LLM + skill

​Tier 2 — LangGraph workflow

​Tier 3 — Full agent with tools and environment

​Tier recommendation by feature

​How tiers interact

​Related