Databricks Agent Deploy Model Serving Dab

Databricks Agent Deploy to Model Serving via DAB

Deploy AI agents to Databricks Model Serving using Databricks Asset Bundles (DAB) for infrastructure-as-code deployment with environment management, MLflow integration, and optional evaluation.

Key Concepts

Deployment Flow

1. Log agent to MLflow with dependencies
2. Register model in Unity Catalog
3. Create DAB configuration with model serving endpoint
4. Deploy using DAB with specified profile
5. Optional: Run evaluation using DAB jobs

Supported Frameworks

• LangGraph: Full LangGraph applications with state management
• OpenAI SDK: Agents using OpenAI Assistants API
• Custom: Any agent following MLflow pyfunc pattern

See references/agent_frameworks.md for complete implementation examples of all frameworks.

Step-by-Step Workflow

Step 1: Analyze Agent Code

Determine: framework type, dependencies, environment variables/secrets, model endpoints, vector search indexes, and tools used.

Step 2: Log Agent to MLflow

Use scripts/log_and_register.py or log manually:

python scripts/log_and_register.py \
    --agent-path src/agent/agent.py \
    --model-name main.agents.my_agent \
    --agent-type langgraph

Step 3: Register Model

model_name = "main.agents.my_agent"  # catalog.schema.model_name
registered_model = mlflow.register_model(model_uri=model_uri, name=model_name)

Step 4: Create DAB Structure

agent-dab/
├── databricks.yml              # Main DAB configuration
├── resources/
│   ├── model_serving.yml       # Model serving endpoint config
│   └── evaluation_job.yml      # Optional: Evaluation job
├── src/agent/                  # Agent implementation
├── scripts/
│   ├── deploy.sh               # Deployment script
│   └── log_and_register.py     # MLflow logging script
└── requirements.txt

Step 5: Generate databricks.yml

bundle:
  name: my-agent-deployment

include:
  - resources/*.yml

variables:
  catalog:
    description: Unity Catalog name
    default: main
  schema:
    default: agents
  model_name:
    default: my_agent
  endpoint_name:
    default: my-agent-endpoint

targets:
  dev:
    mode: development
    default: true
    workspace:
      host: ${DATABRICKS_HOST}
    variables:
      catalog: dev
      endpoint_name: my-agent-dev
  prod:
    mode: production
    workspace:
      host: ${DATABRICKS_HOST}
    variables:
      catalog: prod
      endpoint_name: my-agent-prod
    permissions:
      - level: CAN_MANAGE
        group_name: ml-engineers

Step 6: Generate Model Serving Configuration

Create resources/model_serving.yml:

resources:
  model_serving_endpoints:
    ${var.endpoint_name}:
      config:
        served_entities:
          - entity_name: ${var.catalog}.${var.schema}.${var.model_name}
            entity_version: "1"
            workload_size: Small
            scale_to_zero_enabled: true
            environment_vars:
              DATABRICKS_HOST: ${workspace.host}
        auto_capture_config:
          catalog_name: ${var.catalog}
          schema_name: ${var.schema}
          table_name_prefix: ${var.endpoint_name}
          enabled: true
      permissions:
        - level: CAN_QUERY
          group_name: ml-users
        - level: CAN_MANAGE
          group_name: ml-engineers

Step 7: Deploy

# Using the deployment script
./scripts/deploy.sh dev my-databricks-profile

# Or directly
databricks bundle validate -t dev
databricks bundle deploy -t dev
databricks serving-endpoints query-endpoint \
    --name my-agent-dev \
    --json '{"messages": [{"role": "user", "content": "Hello"}]}'

See references/deployment_commands.md for complete CLI reference.

Best Practices

1. Version control all DAB configurations
2. Separate environments using different catalogs/schemas for dev/staging/prod
3. Secrets management: Use Databricks Secrets for API keys
4. Model versions: Pin versions in production, use "latest" in dev
5. Auto-capture: Enable for monitoring and debugging
6. Testing: Test in dev before promoting to prod

Common Issues

References

• references/agent_frameworks.md - Framework-specific examples (LangGraph, OpenAI, custom RAG)
• references/dab_structure.md - Detailed DAB file structure
• references/deployment_commands.md - CLI command reference
• scripts/deploy.sh - Deployment script
• scripts/log_and_register.py - MLflow logging script