Agent Manifest Specification

Overview

The Agent Manifest is a YAML document that defines an agent's identity, model configuration, tool access, governance policies, evaluation criteria, and Azure AI Foundry deployment parameters. It follows a Kubernetes-inspired structure with apiVersion, kind, metadata, and spec sections. Manifests are deployed as Foundry Agent Applications — first-class ARM resources with versioned deployments, Entra ID RBAC, and private networking.

Schema

The full JSON Schema is available at agent-templates/agent-manifest-schema.json.

Top-Level Fields

Field	Type	Required	Description
`apiVersion`	string	Yes	Schema version (`foundry.azure.com/v1`)
`kind`	string	Yes	Always `AgentManifest`
`metadata`	object	Yes	Agent identity and labels
`spec`	object	Yes	Agent configuration

Metadata

Field	Type	Description
`name`	string	Unique agent identifier
`version`	string	SemVer version string
`labels.class`	string	Agent class: retrieval, action, governance, orchestrator, analysis
`labels.tier`	string	Deployment tier: sandbox, development, staging, production
`labels.owner`	string	Team or individual responsible

Spec — Model Configuration

spec:
  model:
    deployment: gpt-4o          # Model deployment name
    temperature: 0.3            # Sampling temperature
    max_tokens: 4096            # Max response tokens
    top_p: 0.95                 # Nucleus sampling
    response_format: text       # text | json_object

Spec — Tools

spec:
  tools:
    - name: azure_ai_search
      type: azure_ai_search
      connection: ai-search-connection
    - name: file_search
      type: file_search
    - name: code_interpreter
      type: code_interpreter
    - name: servicenow
      type: mcp
      server_url: https://servicenow-mcp.azurewebsites.net

Spec — Governance

spec:
  governance:
    policy_pack: default        # Policy pack reference
    certification_level: standard
    data_boundary: us-east      # Geographic constraint
    escalation_policy: auto     # auto | manual | block
    memory_retention: 30d       # Memory TTL

Spec — Evaluation

spec:
  evaluation:
    pack: builtin               # Eval pack reference
    threshold: 0.8              # Minimum trust score
    schedule: "0 */6 * * *"     # Cron expression
    dataset: datasets/my-domain.jsonl

Spec — Guardrails (v2.0.0)

spec:
  guardrails:
    content_safety: true         # Azure AI Content Safety
    pii_detection: mask          # off | detect | mask
    prompt_attack_defense: true  # Prompt injection protection
    task_adherence: true         # Stay within defined scope
    custom_blocklists:
      - profanity-list
      - competitor-names

Spec — Memory (v2.0.0)

spec:
  memory:
    chat_summary: true           # Azure Agent Memory (Preview)
    user_profile: true           # Per-user context persistence
    isolation: user_scoped       # user_scoped | session_scoped

Spec — Identity & Networking

spec:
  identity:
    type: managed_identity       # Entra ID managed identity
    client_id: <managed-id>
  networking:
    private_endpoint: true       # Azure Private Link
    vnet_injection: true         # VNet integration
    allowed_ip_ranges: []        # IP allowlisting

Validation

Manifests are validated against the JSON Schema at deployment time. The SDK client and CI pipeline both run schema validation before proceeding with Azure AI Foundry agent creation. Validation checks include Azure region availability, model deployment existence, and Entra ID role assignments.