MCP Servers, Memory, Prompt Versions and CI/CD

What it is

A production agent needs more than a model. It reaches tools through MCP servers, keeps state between turns in a persistent datastore, loads its prompts from a versioned Prompt Registry, and moves from development to production through a tested CI/CD process.

Each of these keeps a moving part outside the agent code, so it can change, be governed and be rolled back on its own.

How it works

MCP servers

The Model Context Protocol gives agents a standard way to use tools. An MCP server lists its tools with input schemas, and the MCP client in the agent discovers them at run time, so you do not hard code each tool.

Managed servers are addressed by workspace URLs that name the resource, for example /api/2.0/mcp/genie/{genie_space_id}, and they enforce Unity Catalog permissions. External connections keep credentials in Unity Catalog rather than in agent code.

Persistent memory

Model Serving endpoints are stateless, so memory kept in a Python variable is lost between requests and replicas. A low latency transactional store such as Lakebase fits per turn reads and writes; Delta tables fit analytical or batch reuse of results.

Prompt version control

The MLflow Prompt Registry stores prompts in Unity Catalog with a catalog.schema.prompt name. Each registration creates an immutable version; an alias is a movable pointer to one version.

p = mlflow.genai.register_prompt(
    name="main.support.answer_prompt",
    template="Using only the context, answer: {{question}}",
    commit_message="Restrict answers to context")
mlflow.genai.set_prompt_alias(name="main.support.answer_prompt", alias="production", version=p.version)
prompt = mlflow.genai.load_prompt("prompts:/main.support.answer_prompt@production")
text = prompt.format(question="How do I reset my password?")

Because the agent loads the prompt by alias, moving production to a new version changes behavior without redeploying code, and moving it back is an instant rollback.

CI/CD for agents

• Test components on their own. Check that the retriever returns the expected chunks for known queries, that each tool returns valid output for sample arguments, and that the prompt formats correctly, before testing the whole agent.
• Gate on evaluation. Run the agent against an evaluation dataset and promote only when scorer results meet the bar.
• Promote prompts across environments by registering the approved version in the target environment and moving its alias, never by editing prompt text in production.
• Update the Vector Search index as part of the pipeline: a job writes new chunks to the source Delta table, then calls index.sync() for a triggered Delta Sync index, or upserts for a Direct Vector Access index.
• Deploy as code with Databricks Asset Bundles (called Declarative Automation Bundles in newer docs), so jobs, endpoints and permissions are defined in files and versioned in Git.

How the exam asks about this

• "The agent must query several Genie spaces and Unity Catalog functions without anyone hosting a server." Use managed MCP servers.
• "Users expect the assistant to remember their preferred region in next week's conversation." Store it as long term memory in a persistent store such as Lakebase, not in the endpoint process.
• "A prompt change hurt answer quality in production and must be reverted at once." Move the production alias back to the previous prompt version.