Tiered Enforcement Protocol

Not every task needs the full multi-agent orchestra. This protocol helps match task complexity to the right tooling - balancing capability against overhead (tokens, latency, complexity).

Version: 1.0.0 Status: Mandatory for SPINE-orchestrated projects

---

The Three Tiers

Tier	Task Type	Enforcement	Examples
Tier 1	Simple	None needed	Fix typo, add docstring
Tier 2	Medium	Recommended	New API endpoint, bug investigation
Tier 3	Complex	Mandatory	System design, dashboard implementation

---

Tier 1: Simple Tasks

Single-file edits, typo fixes, small bug fixes, adding comments, simple refactoring within one file, quick questions about code.

What to do: Just do it. No subagents, no MCP, no documentation required.

Examples:

• “Fix the typo in README.md”
• “Add a docstring to this function”
• “Rename this variable”
• “What does this function do?”

Cost: ~500-2,000 tokens, 2-5 seconds, 1 API call

---

Tier 2: Medium Tasks

Multi-file changes (2-5 files), new features with clear scope, bug fixes that need investigation, refactoring across files, integration work.

What to do:

• Use Task(subagent_type="Explore") for codebase exploration
• Use browser-mcp for UI changes
• Use next-conductor for task status
• If you skip any of these, document why

Examples:

• “Add a new API endpoint with tests”
• “Implement dark mode toggle”
• “Fix the authentication bug”
• “Refactor the utils module”

Skip template:

SPINE Usage (Tier 2):
- Subagents: [Used/Skipped - reason]
- browser-mcp: [Used/Skipped - reason]
- next-conductor: [Used/Skipped - reason]

Cost: ~5,000-15,000 tokens, 30-90 seconds, 2-5 API calls

---

Tier 3: Complex Tasks

Architecture decisions, research tasks, UI-heavy implementations, multi-component features (6+ files), system design, performance optimization.

What to do:

• Use the right subagents: code-architect for design, research-coordinator for research, visual-tester + browser-mcp for UI
• Take screenshots for visual verification
• Document capability usage
• Use SPINE Python logging when it makes sense

Examples:

• “Design the authentication system architecture”
• “Research caching best practices for this stack”
• “Implement the dashboard with charts and filters”
• “Optimize database query performance”

Documentation template:

SPINE Usage (Tier 3):
- Subagents used:
  - [subagent_type]: [purpose]
- MCP servers used:
  - [server]: [purpose]
- Screenshots: [location]
- SPINE logs: [experiment ID]

Cost: ~20,000-100,000+ tokens, 2-10 minutes, 5-20+ API calls

---

Why Bother with Tiers?

The cost of over-engineering

Factor	Impact
Token cost	Parallel subagents = 2-6x cost increase
Latency	Subagent spawn adds 10-30 seconds
Over-engineering	Simple tasks don’t need orchestration
Context fragmentation	Subagents don’t share full conversation context
Debugging	More layers = harder to trace failures

The cost of under-engineering

Factor	Impact
Missed requirements	Incomplete exploration
Rework	Architectural mistakes caught late
Quality issues	Insufficient verification
Technical debt	Quick fixes that don’t scale

---

Decision Flow

flowchart TD
    START(["New Task"]) --> Q1{"1 file?"}
    Q1 -- "Yes" --> T1["TIER 1"]
    Q1 -- "No (2-5 files)" --> Q2{"Design decisions<br/>needed?"}
    Q2 -- "Yes" --> T3a["TIER 3"]
    Q2 -- "No" --> Q3{"Research / UI-heavy<br/>/ architecture?"}
    Q3 -- "Yes" --> T3b["TIER 3"]
    Q3 -- "No" --> T2["TIER 2"]

    style START fill:#f59e0b,stroke:#d97706,color:#000
    style Q1 fill:#1e293b,stroke:#475569,color:#e2e8f0
    style Q2 fill:#1e293b,stroke:#475569,color:#e2e8f0
    style Q3 fill:#1e293b,stroke:#475569,color:#e2e8f0
    style T1 fill:#0d9488,stroke:#0d9488,color:#fff
    style T2 fill:#2563eb,stroke:#2563eb,color:#fff
    style T3a fill:#7c3aed,stroke:#7c3aed,color:#fff
    style T3b fill:#7c3aed,stroke:#7c3aed,color:#fff

---

Try It Out

The Tier Classifier Demo walks through the classification with four questions:

1. Scope - how many files/systems affected?
2. Clarity - how clear are the requirements?
3. Decisions - what design decisions needed?
4. Verification - what verification required?

---

Checklists

Before starting (all tiers)

• Classify task tier
• If Tier 2/3: identify which capabilities apply

Tier 2 completion

• If subagents skipped: documented reason
• If browser-mcp skipped for UI: documented reason

Tier 3 completion

• Appropriate subagents used
• browser-mcp used for UI verification
• Screenshots captured
• Capability usage documented

---

Examples: Right vs Wrong

UI Feature

Task: “Add a settings page with theme toggle”

Wrong:

- Created settings.tsx
- Added theme toggle
- Done

Right:

- Used Explore to find existing theme patterns
- Created settings.tsx
- Used browser-mcp to navigate and screenshot
- Used visual-tester to verify rendering

SPINE Usage (Tier 3):
- Subagents: Explore (found ThemeContext), visual-tester (verified UI)
- MCP: browser-mcp (screenshots)
- Screenshots: screenshots/settings-theme-toggle.png

Research Task

Task: “Research caching strategies for our API”

Wrong:

- Read some articles
- Recommended Redis
- Done

Right:

- Used research-coordinator to investigate:
  - Current caching patterns in codebase
  - Industry best practices
  - Framework-specific recommendations
- Synthesized findings with trade-offs

SPINE Usage (Tier 3):
- Subagents: research-coordinator, code-architect
- MCP: research-agent-mcp

Simple Bug Fix

Task: “Fix null pointer in getUserName()”

Fine:

- Found the issue: missing null check
- Added null check with fallback
- Done

No extra documentation needed for Tier 1.

---

How Enforcement Works

This is instruction-based. Claude reads the protocol, classifies tasks, follows requirements, and documents usage. There’s no automated gate - that may come later.

---

Related Docs

• Architecture Overview - system design and components
• Pattern Guide - fan-out and pipeline usage

← Back to Main