ms-ai-architect/skills/ms-ai-advisor/references/development/agent-framework.md

# Microsoft Agent Framework - Knowledge Base

**Last updated:** 2026-01
**Status:** GA (General Availability)

---

## Hva er Microsoft Agent Framework?

Microsoft Agent Framework er Microsofts SDK for å bygge AI-agenter i kode. Det er etterfølgeren til Semantic Kernel og tilbyr et unified rammeverk for agent-utvikling på tvers av Azure AI Foundry og standalone-applikasjoner.

**Nøkkelegenskaper:**
- Multi-agent orkestrering
- Tool/function calling
- Memory og state management
- Streaming og async support
- Azure AI Foundry-integrasjon

**Språk:** Python, C#, JavaScript/TypeScript

---

## Forhold til Semantic Kernel

| Aspekt | Semantic Kernel | Microsoft Agent Framework |
|--------|-----------------|---------------------------|
| **Status** | Vedlikeholdes fortsatt | Anbefalt for nye prosjekter |
| **Fokus** | LLM-orkestrering, plugins | Multi-agent systemer |
| **Abstraksjonsnivå** | Høy | Middels |
| **Azure-integrasjon** | God | Tight (Foundry-native) |
| **Memory** | Basic | Avansert (persistent) |

**Anbefaling:** Bruk Microsoft Agent Framework for nye prosjekter. Semantic Kernel-kode kan migreres gradvis.

---

## Kjernekomponenter

### Agent

En autonom enhet som kan:
- Motta instruksjoner
- Bruke verktøy (tools)
- Samarbeide med andre agenter
- Opprettholde tilstand

```python
from azure.ai.agent import Agent, AgentConfig

agent = Agent(
    config=AgentConfig(
        name="ResearchAgent",
        instructions="Du er en forskningsassistent som finner fakta.",
        model="gpt-4o",
        tools=[search_tool, file_reader_tool]
    )
)
```

### Tools

Funksjoner agenten kan kalle:

```python
from azure.ai.agent import tool

@tool
def search_web(query: str) -> str:
    """Søk på nettet etter informasjon."""
    # Implementasjon
    return results

@tool
def read_file(path: str) -> str:
    """Les innholdet i en fil."""
    # Implementasjon
    return content
```

### Memory

Lagre og hente kontekst på tvers av samtaler:

```python
from azure.ai.agent import Memory

memory = Memory(
    type="persistent",  # eller "session"
    storage="cosmos_db"  # eller "in_memory"
)

agent = Agent(
    config=config,
    memory=memory
)
```

### Multi-Agent Orchestration

Koordiner flere agenter:

```python
from azure.ai.agent import Swarm, Handoff

research_agent = Agent(name="Researcher", ...)
writer_agent = Agent(name="Writer", ...)

swarm = Swarm(
    agents=[research_agent, writer_agent],
    handoffs=[
        Handoff(
            from_agent="Researcher",
            to_agent="Writer",
            condition="research_complete"
        )
    ]
)

result = await swarm.run("Skriv en rapport om AI-trender")
```

---

## Azure AI Foundry-integrasjon

Agent Framework er native integrert med Azure AI Foundry Agent Service.

### Deploye til Foundry

```python
from azure.ai.foundry import FoundryClient

client = FoundryClient(
    endpoint="https://<workspace>.api.azureml.ms",
    credential=DefaultAzureCredential()
)

# Deploye agent
deployment = client.agents.deploy(
    agent=my_agent,
    name="production-agent",
    scaling={
        "min_instances": 1,
        "max_instances": 10
    }
)
```

### Bruke Foundry Tools

Tilgang til 1,400+ Logic Apps connectors:

```python
from azure.ai.foundry import FoundryTools

tools = FoundryTools(client)

# Legg til SharePoint-tilgang
sharepoint = tools.get("sharepoint")
my_agent.add_tool(sharepoint)

# Legg til Fabric-tilgang
fabric = tools.get("fabric")
my_agent.add_tool(fabric)
```

---

## Patterns

### Pattern 1: RAG Agent

```python
from azure.ai.agent import Agent, tool
from azure.ai.search import SearchClient

search_client = SearchClient(...)

@tool
def search_documents(query: str) -> str:
    """Søk i kunnskapsbasen."""
    results = search_client.search(query, top=5)
    return "\n".join([r.content for r in results])

rag_agent = Agent(
    name="KnowledgeAgent",
    instructions="""
    Du er en kunnskapsassistent. Bruk search_documents for å finne
    relevant informasjon før du svarer. Siter alltid kilder.
    """,
    tools=[search_documents]
)
```

### Pattern 2: Supervisor-Worker

```python
from azure.ai.agent import Agent, Swarm

# Worker agents
researcher = Agent(name="Researcher", instructions="Finn fakta...")
writer = Agent(name="Writer", instructions="Skriv innhold...")
reviewer = Agent(name="Reviewer", instructions="Kvalitetssjekk...")

# Supervisor
supervisor = Agent(
    name="Supervisor",
    instructions="""
    Du koordinerer arbeidet mellom Researcher, Writer og Reviewer.
    1. Gi Researcher en research-oppgave
    2. Gi Writer output fra Researcher
    3. La Reviewer validere
    4. Iterer hvis nødvendig
    """,
    sub_agents=[researcher, writer, reviewer]
)
```

### Pattern 3: Human-in-the-Loop

```python
from azure.ai.agent import Agent, Checkpoint

@checkpoint
async def approve_action(action: str) -> bool:
    """Krever menneskelig godkjenning."""
    approval = await request_human_approval(action)
    return approval.approved

agent = Agent(
    name="ActionAgent",
    instructions="Utfør handlinger, men be om godkjenning først.",
    checkpoints=[approve_action]
)
```

### Pattern 4: Streaming Response

```python
from azure.ai.agent import Agent

agent = Agent(...)

# Streaming for responsiv UI
async for chunk in agent.run_stream("Forklar kvantefysikk"):
    print(chunk.text, end="", flush=True)
```

---

## Memory Strategies

### In-Memory (Session)

```python
memory = Memory(type="session")
# Varer kun for denne sesjonen
# Raskest, men ingen persistens
```

### Cosmos DB (Persistent)

```python
memory = Memory(
    type="persistent",
    storage="cosmos_db",
    connection_string="...",
    database="agents",
    container="conversations"
)
# Persisterer på tvers av sesjoner
# Støtter vector search for semantic retrieval
```

### Redis (Distributed)

```python
memory = Memory(
    type="distributed",
    storage="redis",
    connection_string="..."
)
# For multi-instance deployment
# Lavere latency enn Cosmos DB
```

---

## Observability

### Tracing

```python
from azure.ai.agent import enable_tracing
from opentelemetry.sdk.trace.export import ConsoleSpanExporter

enable_tracing(
    exporter=ConsoleSpanExporter(),
    # eller: AzureMonitorExporter()
)

# Alle agent-operasjoner logges nå
```

### Metrics

```python
from azure.ai.agent import metrics

# Agent-level metrics
agent.on_run_complete(lambda m: log_metrics(m))

# Metrics inkluderer:
# - Token usage
# - Tool calls
# - Latency
# - Error rates
```

### Azure Monitor Integration

```python
from azure.monitor.opentelemetry import configure_azure_monitor

configure_azure_monitor(
    connection_string="InstrumentationKey=..."
)

# All telemetry -> Application Insights
```

---

## Security

### Managed Identity

```python
from azure.identity import DefaultAzureCredential

agent = Agent(
    credential=DefaultAzureCredential(),
    # Ingen secrets i koden
)
```

### Content Safety

```python
from azure.ai.contentsafety import ContentSafetyClient

safety = ContentSafetyClient(...)

@tool
def safe_generate(prompt: str) -> str:
    # Sjekk input
    input_check = safety.analyze_text(prompt)
    if input_check.harmful:
        raise ValueError("Harmful input detected")

    # Generer
    response = llm.generate(prompt)

    # Sjekk output
    output_check = safety.analyze_text(response)
    if output_check.harmful:
        return "Kunne ikke generere trygt svar"

    return response
```

### Tool Permission Scoping

```python
@tool(
    permissions=["files.read"],  # Begrensede permissions
    require_confirmation=True     # Krev bekreftelse
)
def read_sensitive_file(path: str) -> str:
    ...
```

---

## Migration fra Semantic Kernel

### Kernel → Agent

```python
# Semantic Kernel
kernel = Kernel()
kernel.add_plugin(MyPlugin())
result = await kernel.invoke(function, input)

# Agent Framework
agent = Agent(tools=[my_tool])
result = await agent.run(input)
```

### Plugins → Tools

```python
# Semantic Kernel plugin
@kernel_function
def my_function(input: str) -> str:
    return process(input)

# Agent Framework tool
@tool
def my_function(input: str) -> str:
    return process(input)
```

### Planners → Orchestration

```python
# Semantic Kernel planner
planner = SequentialPlanner(kernel)
plan = await planner.create_plan(goal)
result = await plan.invoke()

# Agent Framework
agent = Agent(
    instructions=goal,
    tools=[...]
)
result = await agent.run()  # Automatisk planning
```

---

## For Cosmo: Beslutningsveiledning

### Når anbefale Agent Framework

1. **Utviklerteam** som bygger AI-applikasjoner
2. **Multi-agent systemer** med kompleks orkestrering
3. **Tight Azure-integrasjon** via Foundry Agent Service
4. **Custom logic** som krever kode
5. **Produksjonskrav** (observability, scaling, security)

### Når anbefale Copilot Studio istedenfor

1. **Citizen developers** uten kodeerfaring
2. **Rask prototyping** av chatbots
3. **Standard scenarios** (Q&A, IT helpdesk)
4. **Power Platform-økosystem** allerede i bruk

### Når anbefale direkte Azure OpenAI istedenfor

1. **Enkle API-kall** uten orkestrering
2. **Minimal kompleksitet** påkrevd
3. **Eksisterende SDK-integrasjon** (OpenAI SDK)

### Spørsmål å stille kunden

1. "Har dere utviklere som kan skrive Python/C#/TypeScript?"
2. "Trenger dere at flere agenter samarbeider?"
3. "Hvilke systemer må agenten integrere med?"
4. "Hva er kravene til observability og logging?"
5. "Skal løsningen kjøre i Azure, on-prem, eller hybrid?"

---

## Ressurser

- [Agent Framework Documentation](https://learn.microsoft.com/azure/ai-services/agents)
- [Azure AI Foundry Agent Service](https://learn.microsoft.com/azure/ai-foundry/agent-service)
- [Migration Guide from Semantic Kernel](https://learn.microsoft.com/azure/ai-services/agents/migrate-semantic-kernel)
- [GitHub Samples](https://github.com/azure-samples/ai-agent-framework)

---

*Sist oppdatert: Januar 2026*