Migrate From Model Serving

Model Serving to Databricks Apps Migration Guide

This guide instructs LLM coding agents how to migrate an MLflow ResponsesAgent from Databricks Model Serving to Databricks Apps.

Overview

Goal: Migrate an agent deployed on Databricks Model Serving (using ResponsesAgent with predict()/predict_stream()) to Databricks Apps (using MLflow GenAI Server with @invoke/@stream decorators).

Key Transformation:

• Model Serving: Synchronous predict() and predict_stream() methods on a class
• Apps: Functions with @invoke and @stream decorators (sync or async, based on user preference)

Deliverables: After migration is complete, you will have:

<working-directory>/
├── original_mlflow_model/    # Downloaded artifacts from Model Serving
│   ├── MLmodel
│   ├── code/
│   │   └── agent.py
│   ├── input_example.json
│   └── requirements.txt
│
└── <app-name>/               # New Databricks App (ready to deploy)
    ├── agent_server/
    │   ├── agent.py          # Migrated agent code
    │   └── ...
    ├── databricks.yml        # Bundle config with resources
    ├── pyproject.toml
    ├── uv.lock
    └── ...

<app-name> is the name the user provides at the start of the migration. It is used as both the directory name and the Databricks App name at deploy time.

Before You Begin: Gather User Inputs

Before doing anything else, ask the user three questions. Use the AskUserQuestion tool to collect all answers at once so the user is only prompted once, then Claude can execute the rest of the migration autonomously.

Questions to ask:

1. Databricks profile: Which Databricks CLI profile should be used for the workspace where the Model Serving endpoint lives? (Run databricks auth profiles first to list available profiles and their workspaces, then present the options to the user.)
2. App name: What should the new Databricks App be named? (Must be lowercase, can contain letters, numbers, and hyphens, and must be unique within the workspace.)
3. Async migration: Would you like to migrate your agent code to be fully async?
   • Yes (Recommended): Converts all I/O operations to async (await/async for), enabling higher concurrency on smaller compute — no more threads sitting idle while waiting for LLM responses or long-running tool calls.
   • No: Keeps your existing synchronous code with minimal changes — just extracts the logic from the ResponsesAgent class and wraps it with @invoke/@stream decorators. Simpler migration, but each request blocks a thread while waiting for I/O.

Store the answers as:

• <profile> — used for ALL databricks CLI commands throughout the migration (via --profile <profile>)
• <app-name> — used as both the directory name for the migrated app AND the app name when deploying with databricks bundle deploy
• <async> — yes or no, determines whether to convert the agent code to async or keep it synchronous

Validate Authentication

After receiving the user's answers, validate the selected profile:

databricks current-user me --profile <profile>

If this fails with an authentication error, prompt the user to re-authenticate:

databricks auth login --profile <profile>

Important: Remember to include --profile <profile> on every databricks CLI command throughout the migration.

Create the App Directory

Copy all scaffold files from the current working directory into a new directory named <app-name>/. Exclude instruction files (AGENTS.md, CLAUDE.md), hidden directories (.claude/, .git/), and any migration artifacts (e.g., original_mlflow_model/, .migration-venv/). Do NOT search for or copy scaffold files from other directories or templates — everything you need is right here.

All subsequent migration steps operate inside the <app-name>/ directory.

Note: The agent_server/agent.py scaffold is intentionally framework-agnostic — it contains the @invoke/@stream decorator pattern with TODO placeholders. Step 3 (Migrate the Agent Code) will replace these placeholders with the actual agent logic from the original Model Serving endpoint.

Create Task List

Create a task list to track progress. This helps the user follow along and see what's completed, in progress, and pending.

User tip: Press Ctrl+T to toggle the task list view in your terminal. The display shows up to 10 tasks at a time with status indicators.

Create the following tasks using the TaskCreate tool:

Update task status as you progress:

• Mark tasks as in_progress when starting each step
• Mark tasks as completed when finished
• This gives the user visibility into migration progress

Step 1: Download the Original Agent Code

Task: Mark "Authenticate to Databricks" as completed. Mark "Download original agent artifacts" as in_progress.

Note: The <profile> and <app-name> values were collected from the user in the "Before You Begin" section. Use them throughout.

Download the original agent code from the Model Serving endpoint. This requires setting up a virtual environment with MLflow to access the model artifacts.

1.1 Get Model Info from Endpoint

If you have a serving endpoint name, extract the model details:

# Get endpoint info (remember to include --profile if using non-default)
databricks serving-endpoints get <endpoint-name> --profile <profile> --output json

Look for served_entities[0].entity_name (model name) and entity_version in the response. Find the entity with 100% traffic in traffic_config.routes.

1.2 Download Model Artifacts

Use uv run --with to download artifacts without creating a separate virtual environment. The mlflow[databricks] extra includes boto3 for Unity Catalog artifact access:

DATABRICKS_CONFIG_PROFILE=<profile> uv run --no-project \
  --with "mlflow[databricks]>=2.15.0" \
  --with "databricks-sdk>=0.30.0" \
  python3 << 'EOF'
import mlflow

mlflow.set_tracking_uri("databricks")

# Replace with actual values from step 1.1
MODEL_NAME = "<model-name>"
VERSION = "<version>"

print(f"Downloading model: models:/{MODEL_NAME}/{VERSION}")
mlflow.artifacts.download_artifacts(
    artifact_uri=f"models:/{MODEL_NAME}/{VERSION}",
    dst_path="./original_mlflow_model"
)
print("Download complete! Artifacts saved to ./original_mlflow_model")
EOF

1.3 Verify Downloaded Artifacts

Check that the key files exist and understand the full structure:

# List all downloaded files recursively
find ./original_mlflow_model -type f | head -50

# Check for MLmodel file (contains resource requirements)
cat ./original_mlflow_model/MLmodel

# Check for input example (useful for testing)
cat ./original_mlflow_model/input_example.json 2>/dev/null

Examine the /code folder - contains all code dependencies logged via code_paths=["..."]:

# List all code files
ls -la ./original_mlflow_model/code/

# The main agent is typically agent.py, but there may be additional modules
find ./original_mlflow_model/code -name "*.py" -type f

Examine the /artifacts folder (if present) - contains artifacts logged via artifacts={...}:

# Check for artifacts folder
ls -la ./original_mlflow_model/artifacts/ 2>/dev/null

# List all artifacts
find ./original_mlflow_model/artifacts -type f 2>/dev/null

Important: Take note of ALL files in /code and /artifacts. You will need to copy these to the migrated app and ensure imports still work correctly.

Expected Output Structure

After successful download, you should have:

./original_mlflow_model/
├── MLmodel              # Model metadata and resource requirements
├── code/                # Code logged via code_paths=["..."]
│   ├── agent.py         # Main agent implementation
│   ├── utils.py         # (optional) Helper modules
│   ├── tools.py         # (optional) Custom tool definitions
│   └── ...              # Any other code dependencies
├── artifacts/           # (optional) Artifacts logged via artifacts={...}
│   ├── config.yaml      # (optional) Configuration files
│   ├── prompts/         # (optional) Prompt templates
│   └── ...              # Any other artifacts (data files, etc.)
├── input_example.json   # Sample request for testing
├── requirements.txt     # Original dependencies
└── ...

Key Files to Examine

1. code/agent.py - Contains the ResponsesAgent class with predict() and predict_stream() methods
2. code/*.py - Any additional Python modules the agent imports
3. MLmodel - Contains the resources section listing required Databricks resources
4. artifacts/ - Any configuration files, prompts, or data files the agent uses
5. input_example.json - Use this to test the migrated agent

Troubleshooting Model Download

"Unable to import necessary dependencies to access model version files in Unity Catalog" This means boto3 is missing. Ensure you're using mlflow[databricks] (not just mlflow) in the --with flag — the [databricks] extra includes boto3.

"INVALID_PARAMETER_VALUE" or authentication errors Re-authenticate with Databricks (include profile if non-default):

databricks auth login --profile <profile>

Wrong workspace / Model not found Make sure you're using the correct profile that corresponds to the workspace where the model is deployed:

# List profiles to see which workspace each points to
databricks auth profiles

# Verify you can access the workspace
databricks current-user me --profile <profile>

# List models in that workspace
databricks registered-models list --profile <profile>
databricks model-versions list --name "<model-name>" --profile <profile>

Step 2: Understand the Key Transformations

Task: Mark "Download original agent artifacts" as completed. Mark "Analyze and understand agent code" as in_progress.

Entry Point Transformation

In both cases, the ResponsesAgent class is replaced with decorated functions. The difference is whether those functions are async or sync.

Model Serving (OLD):

from mlflow.pyfunc import ResponsesAgent, ResponsesAgentRequest, ResponsesAgentResponse

class MyAgent(ResponsesAgent):
    def predict(self, request: ResponsesAgentRequest, params=None) -> ResponsesAgentResponse:
        # Synchronous implementation
        ...
        return ResponsesAgentResponse(output=outputs)

    def predict_stream(self, request: ResponsesAgentRequest, params=None):
        # Synchronous generator
        for chunk in ...:
            yield ResponsesAgentStreamEvent(...)

Apps — Async (if <async> = yes):

from mlflow.genai.agent_server import invoke, stream
from mlflow.types.responses import (
    ResponsesAgentRequest,
    ResponsesAgentResponse,
    ResponsesAgentStreamEvent,
)

@invoke()
async def non_streaming(request: ResponsesAgentRequest) -> ResponsesAgentResponse:
    # Async implementation - typically calls streaming() and collects results
    outputs = [
        event.item
        async for event in streaming(request)
        if event.type == "response.output_item.done"
    ]
    return ResponsesAgentResponse(output=outputs)

@stream()
async def streaming(request: ResponsesAgentRequest) -> AsyncGenerator[ResponsesAgentStreamEvent, None]:
    # Async generator
    async for event in ...:
        yield event

Apps — Sync (if <async> = no):

from mlflow.genai.agent_server import invoke, stream
from mlflow.types.responses import (
    ResponsesAgentRequest,
    ResponsesAgentResponse,
    ResponsesAgentStreamEvent,
)

@invoke()
def non_streaming(request: ResponsesAgentRequest) -> ResponsesAgentResponse:
    # Same sync logic from original predict(), extracted from the class
    ...
    return ResponsesAgentResponse(output=outputs)

@stream()
def streaming(request: ResponsesAgentRequest):
    # Same sync generator from original predict_stream(), extracted from the class
    for chunk in ...:
        yield ResponsesAgentStreamEvent(...)

Key Differences

Async Patterns (only if <async> = yes)

Skip this section if the user chose synchronous migration. The sync path keeps all original I/O calls as-is.

All I/O operations must be converted to async:

# OLD (sync)
response = client.chat(messages)

# NEW (async)
response = await client.achat(messages)

# OLD (sync iteration)
for chunk in stream:
    yield chunk

# NEW (async iteration)
async for chunk in stream:
    yield chunk

Step 3: Migrate the Agent Code

Task: Mark "Analyze and understand agent code" as completed. Mark "Migrate agent code to Apps format" as in_progress.

3.1 Copy Code Dependencies and Artifacts

The original MLflow model may contain multiple code files and artifacts that need to be migrated.

Copy all code files from /code to agent_server/:

# Copy all Python files from original code folder
cp ./original_mlflow_model/code/*.py ./<app-name>/agent_server/

# If there are subdirectories with code, copy those too
# cp -r ./original_mlflow_model/code/submodule ./<app-name>/agent_server/

Copy artifacts (if present):

# Create an artifacts directory in the migrated app if needed
mkdir -p ./<app-name>/agent_server/artifacts

# Copy all artifacts
cp -r ./original_mlflow_model/artifacts/* ./<app-name>/agent_server/artifacts/ 2>/dev/null || true

Fix import paths after copying:

When code files are moved, imports may break. Check and update imports in all copied files:

# BEFORE (if files were in different locations):
from code.utils import helper_function
from artifacts.prompts import SYSTEM_PROMPT

# AFTER (files are now in agent_server/):
from agent_server.utils import helper_function
# Or if in same directory:
from .utils import helper_function

# For artifacts, update file paths:
# BEFORE:
with open("artifacts/config.yaml") as f:
# AFTER:
import os
config_path = os.path.join(os.path.dirname(__file__), "artifacts", "config.yaml")
with open(config_path) as f:

Important: Review each copied file and ensure all imports resolve correctly. The most common issues are:

• Relative imports that assumed a different directory structure
• Hardcoded file paths to artifacts
• Missing __init__.py files for package imports

3.2 Extract Configuration

From the original agent code, identify and preserve:

• LLM endpoint name (e.g., databricks-claude-sonnet-4-5)
• System prompt
• Tool definitions
• Any custom logic

3.3 Update the Agent Entry Point

The approach depends on whether the user chose async or sync migration.

Path A: Synchronous Migration (<async> = no)

This is the minimal-changes path. Extract the logic from the ResponsesAgent class, wrap it with @invoke/@stream decorators, and keep all code synchronous.

Edit <app-name>/agent_server/agent.py:

1. Replace the scaffold with the original agent logic. The core transformation is extracting the class methods into decorated functions:

from mlflow.genai.agent_server import invoke, stream
from mlflow.types.responses import (
    ResponsesAgentRequest,
    ResponsesAgentResponse,
    ResponsesAgentStreamEvent,
)

# Move any class __init__ or class-level setup to module level
# e.g., client initialization, tool setup, etc.

@invoke()
def non_streaming(request: ResponsesAgentRequest) -> ResponsesAgentResponse:
    # Paste the body of the original predict() method here
    # Remove 'self.' references — replace with module-level variables
    # Remove 'params' parameter (not used in Apps)
    ...
    return ResponsesAgentResponse(output=outputs)

@stream()
def streaming(request: ResponsesAgentRequest):
    # Paste the body of the original predict_stream() method here
    # Remove 'self.' references — replace with module-level variables
    # Remove 'params' parameter (not used in Apps)
    for chunk in ...:
        yield ResponsesAgentStreamEvent(...)

2. Key changes from class to functions:

   • Remove the class MyAgent(ResponsesAgent): wrapper
   • Remove self parameter from all methods
   • Move __init__ logic (client creation, tool setup) to module-level code
   • Replace self.some_attribute with module-level variables
   • Add @invoke() decorator to the non-streaming function
   • Add @stream() decorator to the streaming function

3. Keep all other code as-is — no need to convert sync calls to async, no need to change for to async for, no need to add await.

Path B: Async Migration (<async> = yes)

This path converts all I/O operations to async for higher concurrency. More changes are required, but the result is a more efficient server.

Edit <app-name>/agent_server/agent.py:

1. Update the LLM endpoint:

   LLM_ENDPOINT_NAME = "<your-endpoint-from-original>"
   

2. Update the system prompt:

   SYSTEM_PROMPT = """<your-system-prompt-from-original>"""
   

3. Add your custom tools: If your original agent had custom tools, add them:

   from langchain_core.tools import tool
   
   @tool
   async def my_custom_tool(arg: str) -> str:
       """Tool description."""
       # Your tool logic (make async if needed)
       return result
   

4. Convert all I/O to async:

   • def predict() → async def non_streaming()
   • def predict_stream() → async def streaming()
   • client.chat() → await client.achat()
   • for chunk in stream: → async for chunk in stream:
   • Sync HTTP calls → await async equivalents

5. Preserve any special logic: Migrate any custom preprocessing, postprocessing, or business logic from the original agent.

3.4 Handle Stateful Agents

If original uses checkpointer (short-term memory):

• Add checkpointer with Lakebase integration (use AsyncCheckpointSaver if async, or sync equivalent if sync)
• Configure LAKEBASE_AUTOSCALING_ENDPOINT in .env
• Extract thread_id from request.custom_inputs or request.context.conversation_id

If original uses store (long-term memory):

• Add store with Lakebase integration (use AsyncDatabricksStore if async, or sync equivalent if sync)
• Configure LAKEBASE_AUTOSCALING_ENDPOINT in .env
• Extract user_id from request.custom_inputs or request.context.user_id

Step 4: Set Up the App

Task: Mark "Migrate agent code to Apps format" as completed. Mark "Set up and configure the app" as in_progress.

4.1 Verify Build Configuration

Before installing dependencies, ensure a README file exists (hatchling requires this):

Ensure a README file exists:

# Create a minimal README if one doesn't exist
if [ ! -f "README.md" ]; then
  echo "# Migrated Agent App" > README.md
fi

4.2 Install Dependencies

cd <app-name>
uv sync

4.3 Generate uv.lock for Databricks Apps

Databricks Apps uses uv.lock for reproducible dependency installation. Generate it by running:

uv lock

IMPORTANT: The uv.lock file must be committed to version control. Databricks Apps detects pyproject.toml + uv.lock (with no requirements.txt) and uses uv to install dependencies. If a requirements.txt exists, it takes priority and uv.lock is ignored.

4.4 Run Quickstart

Run the uv run quickstart script to quickly set up your local environment. This is the recommended way to configure the app as it handles all necessary setup automatically.

uv run quickstart

This script will:

1. Verify uv, nvm, and Databricks CLI installations
2. Configure Databricks authentication
3. Configure agent tracing, by creating and linking an MLflow experiment to your app
4. Configure .env with the necessary environment variables

Important: The quickstart script creates the MLflow experiment that the app needs for logging traces and models. This experiment will be added as a resource when deploying the app.

If there are issues with the quickstart script, refer to the manual setup in section 4.5.

4.5 Manual Environment Configuration (Optional)

If you need to manually configure the environment or add additional variables, edit .env:

# Databricks authentication
DATABRICKS_CONFIG_PROFILE=<your-profile>

# MLflow experiment (created by quickstart, or create manually)
MLFLOW_EXPERIMENT_ID=<experiment-id>

# Example: Lakebase for stateful agents
LAKEBASE_AUTOSCALING_ENDPOINT=<your-endpoint>

# Example: Custom API keys
MY_API_KEY=<value>

To manually create an MLflow experiment:

databricks experiments create-experiment "/Users/<your-username>/<app-name>" --profile <profile>

Step 5: Test Locally

Task: Mark "Set up and configure the app" as completed. Mark "Test agent locally" as in_progress.

Test your migrated agent locally before deploying to Databricks Apps. This helps catch configuration issues early and ensures the agent works correctly.

5.1 Start the Server

After the quickstart setup is complete, start the agent server and chat app locally:

cd <app-name>
uv run start-app

Wait for the server to start. You should see output indicating the server is running on http://localhost:8000.

Note: If you only need the API endpoint (without the chat UI), you can run uv run start-server instead.

5.2 Test with Original Input Example

The original model artifacts include an input_example.json file that contains a sample request. Use this to verify your migrated agent produces the same behavior. If there's no valid sample request then figure out a valid sample request to query agent based on its code.

# Check the original input example (from the <app-name> directory)
cat ../original_mlflow_model/input_example.json

Example content:

{"input": [{"role": "user", "content": "What is an LLM agent?"}], "custom_inputs": {"thread_id": "example-thread-123"}}

Test your local server with this input:

# Test with the original input example
curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d "$(cat ../original_mlflow_model/input_example.json)"

5.3 Test Basic Requests

# Non-streaming
curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hello!"}]}'

# Streaming
curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hello!"}], "stream": true}'

5.4 Test with Custom Inputs (for stateful agents)

# With thread_id for short-term memory
curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hi"}], "custom_inputs": {"thread_id": "test-123"}}'

# With user_id for long-term memory
curl -X POST http://localhost:8000/invocations \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hi"}], "custom_inputs": {"user_id": "user@example.com"}}'

5.5 Verify Before Proceeding

Before proceeding to deployment, ensure:

• The server starts without errors
• The original input example returns a valid response
• Streaming responses work correctly
• Custom inputs (thread_id, user_id) are handled properly (if applicable)

Note: Only proceed to Step 6 (Deploy) after confirming the agent works correctly locally.

Step 6: Deploy to Databricks Apps

Task: Mark "Test agent locally" as completed. Mark "Deploy to Databricks Apps" as in_progress.

This step uses Databricks Asset Bundles (DAB) to deploy. The scaffold includes a databricks.yml that you need to update with the app name and resources from the original model.

6.1 Extract Resources from Original Model

The original model's MLmodel file contains a resources section that lists all Databricks resources the agent needs access to. Check ../original_mlflow_model/MLmodel (or ./original_mlflow_model/MLmodel if you're in the parent directory) for content like:

resources:
  api_version: '1'
  databricks:
    lakebase:
    - name: lakebase
    serving_endpoint:
    - name: databricks-claude-sonnet-4-5

6.2 Update databricks.yml with Resources

The scaffold includes a databricks.yml with the experiment resource pre-configured. You need to:

1. Update the app name to <app-name> (the name provided by the user) in both the resources.apps.agent_migration.name field and the targets.prod.resources.apps.agent_migration.name field.
2. Add resources extracted from the original MLmodel file to the resources.apps.agent_migration.resources list.

Resource Type Mapping (MLmodel → databricks.yml):

Note: The experiment resource is already configured in the scaffold databricks.yml and is automatically created by the bundle. You do not need to add it manually.

Example: databricks.yml for an agent with a serving endpoint and UC function:

resources:
  experiments:
    agent_migration_experiment:
      name: /Users/${workspace.current_user.userName}/${bundle.name}-${bundle.target}

  apps:
    agent_migration:
      name: "<app-name>"  # Update to user's app name
      description: "Migrated agent from Model Serving to Databricks Apps"
      source_code_path: ./
      resources:
        - name: 'experiment'
          experiment:
            experiment_id: "${resources.experiments.agent_migration_experiment.id}"
            permission: 'CAN_MANAGE'
        - name: 'serving-endpoint'
          serving_endpoint:
            name: 'databricks-claude-sonnet-4-5'
            permission: 'CAN_QUERY'
        - name: 'python-exec'
          uc_securable:
            securable_full_name: 'system.ai.python_exec'
            securable_type: 'FUNCTION'
            permission: 'EXECUTE'

targets:
  prod:
    resources:
      apps:
        agent_migration:
          name: "<app-name>"  # Same name for production

Example: Adding Lakebase resources (for stateful agents):

        - name: 'postgres'
          postgres:
            branch: "projects/<project-name>/branches/<branch-name>"
            database: "projects/<project-name>/branches/<branch-name>/databases/<database-id>"
            permission: 'CAN_CONNECT_AND_CREATE'

6.3 Deploy with Databricks Asset Bundles

From inside the <app-name> directory, validate, deploy, and run:

# 1. Validate bundle configuration (catches errors before deploy)
databricks bundle validate --profile <profile>

# 2. Deploy the bundle (creates/updates resources, uploads files)
databricks bundle deploy --profile <profile>

# 3. Run the app (starts/restarts with uploaded source code) - REQUIRED!
databricks bundle run agent_migration --profile <profile>

Important: bundle deploy only uploads files and configures resources. bundle run is required to actually start/restart the app with the new code. If you only run deploy, the app will continue running old code!

6.4 Test Deployed App

Task: Mark "Deploy to Databricks Apps" as completed. Mark "Test deployed app" as in_progress.

# Get the app URL
APP_URL=$(databricks apps get <app-name> --profile <profile> --output json | jq -r '.url')

# Get OAuth token
TOKEN=$(databricks auth token --profile <profile> | jq -r .access_token)

# Query the app
curl -X POST ${APP_URL}/invocations \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"input": [{"role": "user", "content": "Hello!"}]}'

Once the deployed app responds successfully:

Task: Mark "Test deployed app" as completed. Migration complete!

6.5 Deployment Troubleshooting

If you encounter issues during deployment, refer to the deploy skill for detailed guidance.

Debug commands:

# Validate bundle configuration
databricks bundle validate --profile <profile>

# View app logs
databricks apps logs <app-name> --profile <profile> --follow

# Check app status
databricks apps get <app-name> --profile <profile> --output json | jq '{app_status, compute_status}'

# Get app URL
databricks apps get <app-name> --profile <profile> --output json | jq -r '.url'

"App already exists" error: If databricks bundle deploy fails because the app already exists, refer to the deploy skill for instructions on binding an existing app to the bundle.

Reference: App File Structure

<app-name>/
├── agent_server/
│   ├── __init__.py
│   ├── agent.py          # Main agent logic - THIS IS WHERE YOU MIGRATE TO
│   ├── start_server.py   # FastAPI server setup
│   ├── utils.py          # Helper utilities
│   └── evaluate_agent.py # Agent evaluation
├── scripts/
│   ├── __init__.py
│   ├── quickstart.py     # Setup script
│   └── start_app.py      # App startup
├── databricks.yml        # Databricks Asset Bundle configuration (resources, config, targets)
├── pyproject.toml        # Dependencies (for local dev with uv)
├── uv.lock               # Lock file for reproducible deploys (must be committed)
├── .env.example          # Environment template
└── README.md

IMPORTANT: The uv.lock file must be committed to version control. Databricks Apps detects pyproject.toml + uv.lock (with no requirements.txt) and uses uv for fully reproducible installs.

Reference: Common Migration Patterns

Pattern 1: Simple Chat Agent

Original:

class ChatAgent(ResponsesAgent):
    def predict(self, request, params=None):
        messages = to_chat_completions_input(request.input)
        response = self.llm.invoke(messages)
        return ResponsesAgentResponse(output=[...])

Migrated (sync):

llm = ...  # Move class-level init to module level

@invoke()
def non_streaming(request: ResponsesAgentRequest) -> ResponsesAgentResponse:
    messages = to_chat_completions_input(request.input)
    response = llm.invoke(messages)
    return ResponsesAgentResponse(output=[...])

@stream()
def streaming(request: ResponsesAgentRequest):
    # Original predict_stream() body, with self. removed
    ...

Migrated (async):

@invoke()
async def non_streaming(request: ResponsesAgentRequest) -> ResponsesAgentResponse:
    outputs = [e.item async for e in streaming(request) if e.type == "response.output_item.done"]
    return ResponsesAgentResponse(output=outputs)

@stream()
async def streaming(request: ResponsesAgentRequest) -> AsyncGenerator[ResponsesAgentStreamEvent, None]:
    messages = {"messages": to_chat_completions_input([i.model_dump() for i in request.input])}
    agent = await init_agent()
    async for event in process_agent_astream_events(agent.astream(messages, stream_mode=["updates", "messages"])):
        yield event

Pattern 2: Agent with Custom Tools

Sync: Keep tools as-is from the original code.

Async: Migrate tools to async LangChain tools:

from langchain_core.tools import tool

@tool
async def search_docs(query: str) -> str:
    """Search the documentation."""
    results = await vector_store.asimilarity_search(query)
    return format_results(results)

Pattern 3: Using LangGraph with create_agent (async only)

from langchain.agents import create_agent
from databricks_langchain import ChatDatabricks

async def init_agent():
    tools = await mcp_client.get_tools()  # MCP tools are async
    model = ChatDatabricks(endpoint=LLM_ENDPOINT_NAME)
    return create_agent(model=model, tools=tools, system_prompt=SYSTEM_PROMPT)

Reference: Useful Resources

• Responses API Docs: https://mlflow.org/docs/latest/genai/serving/responses-agent/
• Agent Framework: https://docs.databricks.com/aws/en/generative-ai/agent-framework/
• Agent Tools: https://docs.databricks.com/aws/en/generative-ai/agent-framework/agent-tool
• databricks-langchain SDK: https://github.com/databricks/databricks-ai-bridge/tree/main/integrations/langchain

Troubleshooting

"Module not found" errors

uv sync  # Reinstall dependencies

Authentication errors

databricks auth login  # Re-authenticate

Lakebase permission errors

• Ensure the Lakebase instance is added as an app resource in Databricks UI
• Grant appropriate permissions on the Lakebase instance

Async errors (async migration only)

• Ensure all I/O calls use async versions (e.g., await client.achat() not client.chat())
• Use async for instead of for when iterating async generators
• If you chose sync migration, these errors should not occur — double-check that you're not mixing sync and async patterns