aiwg/agentic/code/frameworks/sdlc-complete/schemas/flows/moe-agent-routing.yaml

Cognitive architecture for AI-augmented software development with structured memory, ensemble validation, and closed-loop correction. FAIR-aligned artifacts, 84% cost reduction via human-in-the-loop, standards adopted by 100+ organizations.

# Mixture of Experts Agent Routing Schema
# Based on REF-007 Mixture of Experts
# Issues: #190, #191

$schema: "https://json-schema.org/draft/2020-12/schema"
$id: "https://aiwg.io/schemas/moe-agent-routing/v1"
title: "MoE Agent Routing and Sparse Activation Schema"
description: |
  Schema for Mixture of Experts inspired routing and sparse agent activation
  per REF-007 MoE.

type: object
required:
  - version
  - routing_config
  - activation_policy

properties:
  version:
    type: string
    pattern: "^\\d+\\.\\d+\\.\\d+$"
    default: "1.0.0"

  routing_config:
    $ref: "#/$defs/RoutingConfig"

  activation_policy:
    $ref: "#/$defs/ActivationPolicy"

$defs:
  RoutingConfig:
    type: object
    description: "MoE-inspired routing configuration"
    properties:
      enabled:
        type: boolean
        default: true

      top_k:
        type: integer
        default: 3
        description: "Number of agents to activate per task"

      scoring_weights:
        type: object
        properties:
          domain_match:
            type: number
            default: 0.40
          operation_match:
            type: number
            default: 0.30
          complexity_fit:
            type: number
            default: 0.20
          artifact_match:
            type: number
            default: 0.10

      load_balancing:
        type: object
        properties:
          enabled:
            type: boolean
            default: true
          min_utilization:
            type: number
            default: 0.2
          max_utilization:
            type: number
            default: 0.8
          prefer_underutilized:
            type: boolean
            default: true

  ActivationPolicy:
    type: object
    description: "Sparse activation policy"
    properties:
      max_active:
        type: integer
        default: 5
        description: "Maximum agents in memory"

      ttl_seconds:
        type: integer
        default: 300
        description: "Time-to-live for inactive agents"

      preload:
        type: array
        items:
          type: string
        default:
          - orchestrator
        description: "Always-active agents"

      eager_load:
        type: array
        items:
          type: string
        default: []
        description: "Load at session start"

      lazy_load:
        type: boolean
        default: true
        description: "Load agents on first use"

# Task feature extraction
task_features:
  type: object
  properties:
    domain:
      type: string
      enum:
        - requirements
        - architecture
        - implementation
        - testing
        - security
        - deployment
        - documentation
        - general
    operation:
      type: string
      enum:
        - create
        - review
        - analyze
        - refactor
        - test
        - deploy
        - document
        - debug
    complexity:
      type: integer
      minimum: 1
      maximum: 10
    artifacts:
      type: array
      items:
        type: string
      description: "Referenced artifact types"
    context_tags:
      type: array
      items:
        type: string
      description: "Project tags, tech stack"
    word_count:
      type: integer
    has_artifact_references:
      type: boolean
    has_multiple_phases:
      type: boolean

# Agent capability profile
agent_capability:
  type: object
  required:
    - agent_name
    - domains
    - operations
  properties:
    agent_name:
      type: string
    domains:
      type: object
      additionalProperties:
        type: number
        minimum: 0
        maximum: 1
      description: "Domain → expertise score"
    operations:
      type: object
      additionalProperties:
        type: number
        minimum: 0
        maximum: 1
      description: "Operation → proficiency"
    complexity_range:
      type: object
      properties:
        min:
          type: integer
          default: 1
        max:
          type: integer
          default: 10
        optimal:
          type: integer
          default: 5
    artifact_types:
      type: array
      items:
        type: string
    tags:
      type: array
      items:
        type: string

# Routing result schema
routing_result:
  type: object
  required:
    - task
    - selected_agents
  properties:
    task:
      type: string
    features:
      $ref: "#/$defs/task_features"
    confidence:
      type: number
      minimum: 0
      maximum: 1
    reasoning:
      type: string
    selected_agents:
      type: array
      items:
        type: object
        properties:
          agent:
            type: string
          score:
            type: number
          reasoning:
            type: string
    routing_time_ms:
      type: number

# Agent activation state
activation_state:
  type: object
  properties:
    active_agents:
      type: array
      items:
        type: object
        properties:
          agent:
            type: string
          loaded_at:
            type: string
            format: date-time
          last_used:
            type: string
            format: date-time
          task_count:
            type: integer
          memory_mb:
            type: number
    total_active:
      type: integer
    max_allowed:
      type: integer
    preloaded:
      type: array
      items:
        type: string

# Agent protocol
agent_protocol:
  extract_features:
    description: "Extract features from task description"
    steps:
      - tokenize_task
      - identify_domain_keywords
      - identify_operation_keywords
      - estimate_complexity
      - detect_artifact_references
      - extract_context_tags
      - return_features

  route_task:
    description: "Route task to appropriate agents"
    steps:
      - extract_features
      - load_agent_capabilities
      - for_each_agent:
          - calculate_domain_score
          - calculate_operation_score
          - calculate_complexity_fit
          - calculate_artifact_match
          - compute_total_score
      - apply_load_balancing
      - select_top_k_agents
      - return_routing_result

  activate_agents:
    description: "Activate selected agents"
    steps:
      - get_current_active
      - for_each_selected:
          - if_not_loaded:
              - load_agent
              - add_to_active
      - prune_inactive
      - return_activated_agents

  prune_inactive:
    description: "Deactivate stale agents"
    steps:
      - get_current_time
      - for_each_active:
          - if_ttl_exceeded:
              - if_not_preloaded:
                  - unload_agent
                  - remove_from_active

  update_quality_feedback:
    description: "Update routing weights from outcomes"
    triggers:
      - task_completed
    steps:
      - get_task_result
      - get_selected_agents
      - calculate_quality_score
      - update_agent_scores
      - persist_metrics

# CLI commands
cli_commands:
  route_explain:
    command: "aiwg route-explain <task>"
    description: "Show routing decision for task"

  agents_active:
    command: "aiwg agents active"
    description: "Show currently active agents"

  agents_stats:
    command: "aiwg agents stats"
    description: "Show agent utilization statistics"

  route_override:
    command: "aiwg task <task> --agents <list>"
    description: "Override automatic routing"

# Performance targets (from REF-007)
research_targets:
  routing_latency: "<50ms per decision"
  memory_reduction: "60%+ vs full load"
  activation_latency: "<100ms per agent"
  quality_maintenance: "Equal or better vs. all-agents"

# Storage
storage:
  capabilities_path: ".aiwg/agent-selection/capabilities.json"
  metrics_path: ".aiwg/agent-selection/routing-metrics.json"
  activation_log: ".aiwg/logs/agent-activation.jsonl"

# References
references:
  research:
    - "@.aiwg/research/findings/REF-007-mixture-of-experts.md"
  implementation:
    - "#190"
    - "#191"
  related:
    - "@agentic/code/frameworks/sdlc-complete/schemas/flows/uct-agent-selection.yaml"
    - "@agentic/code/frameworks/sdlc-complete/schemas/flows/semantic-agent-discovery.yaml"