Nomenclature & Glossary - Ensemble Edge

Core Terms

Component

What: A versioned artifact stored in Git that can be independently versioned and deployed. Types: Prompts (.md), configs (.json/.yaml), queries (.sql), scripts (.js/.ts), schemas, templates Purpose: Reusable building blocks that agents reference during execution Managed by: Edgit versioning system Example:

edgit tag create extraction-prompt v1.0.0
edgit tag set extraction-prompt prod v1.0.0
edgit push --tags --force

Used in:

agents:
  - name: analyzer
    component: extraction-prompt@v1.0.0

Agent

What: An executable unit that performs tasks using operations and components Types:

Custom agents - Defined by you using operations
Starter Kit agents - Ships with Conductor (scraper, validator, RAG, etc.)

Capabilities:

Use components
Use operations
Access tools (MCP)
Call other agents
Call ensembles

Example:

agents:
  - name: analyzer
    operation: think
    component: prompt@v2.1.0
    config:
      model: claude-3-5-sonnet-20241022

Operation

What: A primitive execution type that defines HOW something runs Purpose: The “verb” of what an agent does - the execution primitive Types:

think - AI reasoning (LLMs, embeddings, classification)
code - JavaScript/TypeScript execution
storage - Database/KV/R2/D1 operations
http - HTTP requests to external services
tools - MCP/skill access
email - Email operations
sms - SMS operations
html - HTML rendering
pdf - PDF generation
page - Full-stack pages

Usage: Operations are used BY agents, not called directly Example:

agents:
  - name: fetcher
    operation: http  # Operation type
    config:
      url: https://api.example.com

Ensemble

What: An orchestration workflow that coordinates multiple agents Format: YAML files in your Git repository Purpose: Defines workflow logic and execution flow Capabilities:

Orchestrate agents
Call other ensembles
Define flow control (sequential, parallel, conditional, loops)
Manage state
Map inputs/outputs

Cannot:

Use operations directly (must go through agents)
Access components directly (must go through agents)

Example:

ensemble: company-intel

agents:
  - name: fetch
    operation: http

  - name: analyze
    operation: think
    component: prompt@v1.0.0

  - name: store
    operation: storage

Note: Ensembles are NOT versioned. They’re like Kubernetes manifests - you update them to point to different versions of agents and components.

Tool

What: External capability accessed via MCP (Model Context Protocol) or other protocols Access: Through the tools operation in agents Examples: GitHub MCP, filesystem MCP, Claude skills, custom MCP servers Usage:

agents:
  - name: processor
    operation: tools
    config:
      mcp_server: github
      tool: create_issue
    input:
      title: "Bug report"
      body: ${analyze.output}

Execution Terms

State

What: Shared data across agents in an ensemble execution Properties:

Immutable (each update creates new version)
Scoped to ensemble execution
Schema-validated
Access-tracked

Usage:

state:
  schema:
    company_data: object
    analysis: object

agents:
  - name: fetch
    operation: http
    state:
      set: [company_data]  # Write

  - name: analyze
    operation: think
    state:
      use: [company_data]  # Read
      set: [analysis]       # Write

Context

What: The runtime environment and data available during agent execution Includes:

Input data
Previous agent outputs
State variables
Environment variables
Request metadata

Accessed via interpolation:

input:
  data: ${fetch.output}           # Previous agent output
  domain: ${input.domain}          # Ensemble input
  company: ${state.company_data}   # State variable
  api_key: ${env.API_KEY}          # Environment variable

Interpolation

What: Template syntax for referencing dynamic values Syntax: ${source.path.to.value} Sources:

input.* - Ensemble input
<agent-name>.output.* - Agent outputs
state.* - State variables
env.* - Environment variables

Examples:

${input.domain}                    # Simple value
${fetch.output.data.company_name}  # Nested object
${state.analyses[0].score}         # Array access
${env.ANTHROPIC_API_KEY}           # Environment

Versioning Terms

Tag

What: A Git tag in the format components/{name}/{version} or agents/{name}/{version} Purpose: Mark specific commits as versions of components or agents Examples:

components/extraction-prompt/v1.0.0
components/extraction-prompt/prod

agents/scraper/v1.5.0
agents/scraper/prod

Deployment

What: The act of marking a specific version as active in an environment Environments: Typically prod, staging, dev, canary (you define these) Example:

edgit tag set extraction-prompt prod v1.0.0
edgit tag set extraction-prompt staging v2.0.0-beta
edgit push --tags --force

Version Multiverse

What: The ability to run multiple versions of components and agents simultaneously Purpose: A/B testing, gradual rollouts, bug reproduction Example:

agents:
  # Stable combination
  - name: prod-path
    agent: analyzer@v1.0.0
    component: prompt@v1.0.0

  # Experimental combination
  - name: exp-path
    agent: analyzer@v2.0.0-beta
    component: prompt@v2.1.0

  # Hybrid combination
  - name: hybrid-path
    agent: analyzer@v1.0.0
    component: prompt@v2.1.0

Platform Terms

Workers

What: Cloudflare Workers - V8 isolates that run your code at the edge Properties:

Sub-50ms cold starts
200+ global locations
Automatic scaling
No server management

KV

What: Cloudflare KV - Distributed key-value storage Use cases:

Caching HTTP responses
Caching agent outputs
Storing deployment metadata

Properties:

Eventually consistent
Global replication
Low latency reads

D1

What: Cloudflare D1 - Distributed SQL database Use cases:

Storing structured data
Analytics and reporting
Audit trails

Properties:

SQLite-compatible
Strongly consistent per location
Global replication

R2

What: Cloudflare R2 - Object storage (S3-compatible) Use cases:

Document storage for RAG
Large file uploads
Backups

Properties:

No egress fees
S3-compatible API
Global availability

Vectorize

What: Cloudflare Vectorize - Vector database for embeddings Use cases:

Semantic search
RAG pipelines
Similarity matching

Properties:

Native integration with Workers AI
Global distribution
Low-latency queries

Durable Objects

What: Cloudflare Durable Objects - Strongly consistent, stateful objects Use cases:

HITL (human-in-the-loop) state
Execution state tracking
Coordination and locking

Properties:

Strongly consistent
Single-threaded per object
Global distribution

AI Gateway

What: Cloudflare AI Gateway - Caching and observability layer for AI providers Benefits:

Persistent caching of LLM responses
Usage analytics and logging
Rate limiting and cost control
Works with OpenAI, Anthropic, Workers AI, etc.

Conductor-Specific Terms

Catalog

What: Starter Kit components and agents that ship with Conductor Location: catalog/ directory in Conductor repository Includes:

Starter Kit agents (scraper, validator, RAG, etc.)
Example components (prompts, queries, configs)
Example ensembles

Starter Kit Agent

What: Production-ready agents that ship with Conductor Available:

scraper - Web scraping with 3-tier fallback
validator - Quality scoring with multiple evaluators
rag - RAG pipeline with R2 storage
hitl - Human-in-the-loop approval workflows
fetcher - HTTP client with retry logic
transformer - Data transformation utilities
scheduler - Delayed/scheduled execution

Usage:

agents:
  - name: scraper
    agent: scraper  # Reference starter kit agent
    config:
      url: https://example.com

Flow Control

What: Mechanisms for controlling agent execution order Types:

Sequential - Default, one after another
Parallel - Multiple agents run simultaneously
Conditional - Run based on conditions
Loops - Iterate over collections
Retry - Automatic retry with backoff

Example:

agents:
  # Sequential (default)
  - name: step1
    operation: http

  - name: step2
    operation: think

  # Parallel
  - parallel:
      - name: analyze-a
        operation: think
      - name: analyze-b
        operation: think

  # Conditional
  - name: manual-review
    agent: hitl
    condition: ${step2.score < 0.8}

  # Loop
  - name: process-batch
    operation: code
    loop:
      items: ${input.companies}

  # Retry
  - name: api-call
    operation: http
    retry:
      max_attempts: 3
      backoff: exponential

Edgit-Specific Terms

Component Registry

What: The .edgit/components.json file that tracks all components Contains:

Component names and types
File paths
Metadata
Dependencies

Managed by: Edgit automatically

Discovery

What: The process of finding potential components in your repository Commands:

edgit discover scan --type prompt      # Find all prompts
edgit discover detect src/function.ts  # Analyze specific file
edgit components list --untracked      # Show unregistered components

Deployment Tag

What: A Git tag marking which version is deployed to an environment Format: components/{name}/{environment} or agents/{name}/{environment} Example:

components/extraction-prompt/prod       v1.0.0
components/extraction-prompt/staging    v2.0.0-beta

Common Patterns

Component Versioning Pattern

# Create version
edgit tag create my-component v1.0.0

# Tag to staging and push
edgit tag set my-component staging v1.0.0
edgit push --tags --force

# Test and promote to production
edgit tag set my-component prod v1.0.0
edgit push --tags --force

Agent Definition Pattern

# Custom agent using operation
agents:
  - name: my-agent
    operation: think
    component: prompt@v1.0.0
    config:
      model: claude-3-5-sonnet-20241022
    input:
      data: ${previous-agent.output}
    condition: ${previous-agent.success}
    retry:
      max_attempts: 3

Ensemble Orchestration Pattern

ensemble: my-workflow
description: Multi-step AI workflow

input:
  domain:
    type: string
    required: true

state:
  schema:
    company_data: object
    analysis: object

agents:
  - name: fetch
    operation: http
    state:
      set: [company_data]

  - name: analyze
    operation: think
    state:
      use: [company_data]
      set: [analysis]

  - name: store
    operation: storage
    state:
      use: [analysis]

output:
  domain: ${input.domain}
  result: ${state.analysis}

Capability Matrix

What each entity can do:

Entity	Use Components	Use Operations	Use Tools	Call Agents	Call Ensembles
Ensemble	❌	❌	❌	✅	✅ (nested)
Agent	✅	✅	✅	✅	✅
Operation	✅	N/A	N/A	❌	❌
Component	N/A	N/A	N/A	N/A	N/A

Key Insight: Agents are the only entity that can use operations and components directly. Ensembles orchestrate agents.

Quick Reference

Common Confusions

Q: Are ensembles versioned? A: Yes! With Edgit, ensembles can be versioned independently using edgit tag create my-ensemble v1.0.0 (type is inferred from file location). They reference versioned agents and components, and the ensemble itself can be version-pinned. Q: Can ensembles use operations directly? A: No. Ensembles orchestrate agents. Agents use operations. Q: What’s the difference between an agent and an operation? A: Operations are primitives (think, code, http). Agents use operations to do work. Q: Can I version scripts? A: Yes. Scripts are components and can be versioned with Edgit. Q: What’s a starter kit agent? A: Production-ready agents that ship with Conductor (scraper, validator, RAG, etc.) Q: Can agents call other agents? A: Yes. Agents can call both other agents and ensembles.

Next Steps

Edgit Documentation

Component versioning system

Conductor Documentation

Edge orchestration framework

Operations Reference

All available operations

Starter Kit

Ready-to-use agents

​Core Terms

​Component

​Agent

​Operation

​Ensemble

​Tool

​Execution Terms

​State

​Context

​Interpolation

​Versioning Terms

​Tag

​Deployment

​Version Multiverse

​Platform Terms

​Workers

​KV

​D1

​R2

​Vectorize

​Durable Objects

​AI Gateway

​Conductor-Specific Terms

​Catalog

​Starter Kit Agent

​Flow Control

​Edgit-Specific Terms

​Component Registry

​Discovery

​Deployment Tag

​Common Patterns

​Component Versioning Pattern

​Agent Definition Pattern

​Ensemble Orchestration Pattern

​Capability Matrix

​Quick Reference

​Common Confusions

​Next Steps

Edgit Documentation

Conductor Documentation

Operations Reference

Starter Kit

Core Terms

Component

Agent

Operation

Ensemble

Tool

Execution Terms

State

Context

Interpolation

Versioning Terms

Tag

Deployment

Version Multiverse

Platform Terms

Workers

KV

D1

R2

Vectorize

Durable Objects

AI Gateway

Conductor-Specific Terms

Catalog

Starter Kit Agent

Flow Control

Edgit-Specific Terms

Component Registry

Discovery

Deployment Tag

Common Patterns

Component Versioning Pattern

Agent Definition Pattern

Ensemble Orchestration Pattern

Capability Matrix

Quick Reference

Common Confusions

Next Steps