Skip to content

AgentDef — Microsoft Azure AI Ecosystem Mapping

AgentDef ↔ Microsoft Copilot Studio ↔ Azure AI Foundry

This document describes how AgentDef concepts map to the main components of:

  • Microsoft Copilot Studio
  • Azure AI Foundry
  • Azure AI Agents
  • Semantic Kernel
  • Prompt Flow
  • Azure OpenAI

The goal is to decouple the semantic definition of the agent from the Microsoft-specific implementation.

Note: for Microsoft 365 Copilot declarative agents BOTH directions are implemented and round-trip tested: agentdef adapt m365copilot (adapter) and agentdef import m365copilot (importer). Copilot Studio has a working, validated importer (agentdef import copilotstudio, 94 real agent docs, 0 failures). The Azure AI Foundry, Semantic Kernel, and Prompt Flow mappings below remain conceptual/proposed; no adapter or importer exists for those yet.

1. High-Level Mapping

AgentDef Concept Copilot Studio Azure AI Foundry Semantic Kernel Prompt Flow
agent.md Copilot Instructions Agent System Prompt Kernel Instructions Flow Context
manifest.yaml Copilot Configuration Agent Definition Kernel Builder Flow Definition
instructions/ Topics + Instructions Prompt Templates Prompt Functions Prompt Nodes
skills/ Actions Tools / Functions Skills / Plugins Flow Components
workflows/ Topics / Orchestration Agent Flows Planners Directed Flows
memory/ Conversation Memory Thread State Memory Connectors State Variables
knowledge/ Dataverse / Knowledge Sources RAG Indexes Memory Stores Retrieval Nodes
tools/ Power Platform Connectors Azure Functions / APIs Plugins Tool Nodes
runtime/ Environment Config Deployment Config Kernel Config Runtime Config
evals/ Test Chats Evaluations Evaluation Pipelines Batch Runs
telemetry/ Analytics Azure Monitor Observability Traces
framework/ Export/Import Deployment Templates Kernel Adapters Flow Export

2. Copilot Studio Mapping

Conceptual Equivalent

Copilot Studio functions primarily as:

  • orchestration layer
  • conversational runtime
  • business workflow interface

It is especially oriented toward:

  • no-code / low-code
  • enterprise copilots
  • Power Platform integration
  • conversational automation

Suggested AgentDef Adapter

framework/
└── microsoft-copilot-studio/
    ├── copilot-instructions.md
    ├── topics/
    ├── actions/
    ├── connectors/
    └── environment.yaml

2.1 agent.md → Copilot Instructions

The main content of:

  • identity
  • role
  • behavior
  • style

becomes:

  • base instructions
  • generative orchestration instructions

Example

You are an editorial AI assistant specialized in summarizing weekly Twitter/X content into concise professional briefings.

2.2 skills/ → Actions

AgentDef

skills/
└── summarization/

Copilot Studio Equivalent

  • Power Platform Actions
  • AI Actions
  • Prompt Actions
  • Connector Actions

Example Actions

AgentDef Skill Copilot Action
summarization AI Builder Prompt
classification Topic Classifier
prioritization Decision Flow
retrieval Connector Action

2.3 workflows/ → Topics

Copilot Studio organizes most logic as:

  • Topics
  • Trigger phrases
  • Conversation branches

Example Mapping

AgentDef Workflow Step Copilot Topic
ingest links Input Topic
summarize content AI Prompt Action
group by topic Classification Branch
generate report Output Topic

2.4 memory/ → Conversation State

Equivalent Components

  • Variables
  • Session state
  • Dataverse records
  • Conversation history

Example

memory/
├── session.md
└── profile.json

could map to:

  • conversation variables
  • user profile tables
  • persistent Dataverse entities

2.5 tools/ → Power Platform Connectors

Examples

AgentDef Tool Copilot Connector
twitter api Custom Connector
sharepoint Native Connector
jira Connector
outlook Connector
teams Connector

3. Azure AI Foundry Mapping

Conceptual Equivalent

Azure AI Foundry acts more like:

  • agent engineering platform
  • orchestration environment
  • AI infrastructure layer

Compared to Copilot Studio:

  • lower-level
  • more programmable
  • more composable
  • more suitable for advanced agents

Suggested AgentDef Adapter

framework/
└── azure-foundry/
    ├── agent.yaml
    ├── prompts/
    ├── tools/
    ├── flows/
    ├── evals/
    └── deployments/

3.1 agent.md → Agent Definition

Maps into:

  • Azure AI Agent instructions
  • system prompts
  • orchestration directives

3.2 skills/ → Agent Tools

Azure Foundry Equivalents

AgentDef Skill Azure Equivalent
summarization Prompt Tool
clustering Python Tool
retrieval RAG Tool
classification Function Tool
reasoning Agent Planner

3.3 workflows/ → Prompt Flow

Prompt Flow is conceptually very close to AgentDef workflows.

Example

workflows/
└── weekly_digest.md

maps to:

promptflow/
├── ingest_node
├── extraction_node
├── clustering_node
├── ranking_node
└── report_node

3.4 knowledge/ → Azure AI Search / RAG

Equivalent Services

AgentDef Knowledge Azure Service
documents Azure Blob Storage
ontology CosmosDB / Graph
embeddings Azure AI Search
retrieval RAG pipeline

3.5 memory/ → Thread State

Equivalent Components

  • Agent Thread Memory
  • CosmosDB
  • Redis
  • Stateful orchestration

3.6 runtime/ → Deployment Configuration

Example Mapping

model: gpt-4.1
temperature: 0.2
max_tokens: 4000

becomes:

  • Azure OpenAI deployment config
  • model routing policy
  • inference profile

3.7 evals/ → Azure Evaluations

Equivalent Components

AgentDef Eval Azure Equivalent
hallucination tests AI Evaluations
regressions Batch Evaluations
golden outputs Benchmark datasets
quality metrics Evaluation pipelines

4. Semantic Kernel Mapping

Semantic Kernel is a particularly good fit for AgentDef concepts.

Suggested Structure

framework/
└── semantic-kernel/
    ├── plugins/
    ├── planners/
    ├── memory/
    └── kernel.yaml

Mapping Table

AgentDef Semantic Kernel
skills plugins
workflows planners
memory memory connectors
tools native functions
instructions semantic functions
orchestration kernel execution

Important Principle

Do NOT build agents directly around:

  • Copilot Studio
  • Foundry
  • LangGraph
  • OpenAI SDK

Instead:

AgentDef Canonical Definition
            ↓
   Microsoft Adapter Layer
            ↓
Copilot Studio / Foundry / SK

This allows:

  • migration
  • portability
  • multi-runtime deployment
  • framework independence
enterprise-agent/
│
├── agent.md
├── manifest.yaml
│
├── instructions/
├── skills/
├── workflows/
├── memory/
├── knowledge/
├── evals/
│
└── framework/
    ├── copilotstudio/
    │   └── agent-doc.md        # agentdef adapt copilotstudio (planned)
    └── m365copilot/
        └── declarativeAgent.json   # agentdef adapt m365copilot

Keep the canonical definition as the single source of truth and treat everything under framework/ as generated output: declare the targets in a sync: block and let agentdef sync --check fail CI whenever a generated Microsoft-side file drifts from the definition.