Genie agent mode vs chat mode: when to use which, and why you cannot switch mid-chat

A short thread on the Databricks Community forum caught our eye this month. A Databricks employee had posted an explainer on Genie’s two answer modes, chat and agent. One reader then asked a very practical question: if I start in chat mode and my follow-up needs deeper analysis, do I have to open a new chat to use agent mode? A Databricks employee replied that, as far as they could tell, yes. The mode is set when the conversation starts, context does not carry across chats, and they suggested pasting the original question, filters, definitions and generated SQL into a fresh agent-mode chat. They also hedged, because the documentation does not spell it out.

That hedge is the interesting part. Here is what the docs actually say, and how we would design around it.

What the two modes are

According to the Databricks documentation:

Chat mode works with structured data only. It answers questions that a single SQL query can handle. It cannot analyze unstructured files, use MCP connectors, or break a question into steps.
Agent mode (formerly Research Agent) builds a research plan, runs multiple SQL queries, and returns a report with citations, visualizations and supporting tables. It can also query unstructured files in Unity Catalog volumes and use Databricks-provided MCP connectors. It takes longer to respond.
Flash is a faster option inside agent mode, currently in Beta. A workspace admin has to turn it on from the Previews page, and its reports are more concise. For multi-step “why did this change” questions, the docs recommend standard agent mode.

In both modes, generated queries are read-only, and Unity Catalog checks data access against each end user’s identity.

Can you switch mid-conversation?

The docs say you choose Flash or standard agent mode when you start a conversation and that the conversation keeps the choice. To change options, start a new conversation. They do not explicitly say whether an existing chat can move between chat and agent mode. The forum reply is consistent with that: assume you cannot, and plan for it.

We would not build a workflow that depends on switching. Treat the mode as a property of the question, decided up front.

A simple rule for choosing

Ask what shape the answer has.

“What” and “how many” questions: chat mode. Revenue last month, open tickets by region, top ten suppliers. One query, fast, easy to check against the SQL Genie shows.
“Why” and “what’s driving” questions: agent mode. Why did Q3 run over budget? What changed in churn? These need several queries in sequence, and the report format with citations to each step is the point.
Questions that involve documents or external tools: agent mode, since chat mode cannot touch unstructured files or MCP connectors.

Because agent mode is slower, defaulting everyone to it is a mistake. A known lookup gets slower and harder to eyeball, with no gain.

Carrying context into a new chat

When a chat-mode answer raises a “why”, the forum reply’s workaround is sensible, and we would formalize it:

Open a new chat with agent mode selected before the first message.
Paste the original question, the filters you used, and any business definitions that mattered.
Include the SQL Genie generated, so the agent starts from the same logic rather than re-deriving it.

If your users do this often, put it in the training material rather than leaving them to discover it.

Agent mode does not fix a weak data model

The same forum explainer warned that results depend on how well Genie is briefed on terms like “spend” or “budget”. We agree. A multi-step investigation built on ambiguous definitions just produces a longer, more confident-looking wrong answer. Everything that makes chat mode accurate still applies, and matters more:

Describe tables and columns, and declare keys so joins are correct.
Define key measures once, for example in metric views, so “spend” means one thing.
Build a benchmark set and re-run it when you change instructions or add tables. Databricks documents benchmarks for scoring responses against reference SQL.

Note that benchmarks score SQL answers. A narrative agent-mode report still needs a human to spot-check the cited steps, at least until you trust the domain.

Things we would verify in your environment

Availability. The docs point to Databricks Designated Services for regional availability of agent mode, and Flash has its own geography rules (in-Geo in US and EU workspaces; elsewhere only with cross-Geo processing enabled). Check yours before promising it to a business unit.
Preview status. Flash is Beta. The docs we read do not state a release stage for agent mode itself, so confirm in your workspace and with your account team.
Where users will ask. The agent mode page does not mention Slack or Teams, so do not assume the experience matches there. Test the channel your users will actually use.
Latency expectations. Tell users up front that agent mode takes longer, or they will assume it hung.

The short version

Chat mode is for single-query “what” questions; agent mode is for multi-step “why” questions and for files or MCP tools.
The mode is chosen when the conversation starts. Do not design around switching mid-chat.
When you need to escalate, start a new agent-mode chat and paste in the question, filters, definitions and SQL.
Don’t default everyone to agent mode; it is slower.
Agent mode amplifies your data model, good or bad. Fix definitions and benchmarks first.
Check regional availability and release status in your own workspace.

If you are deciding which Genie experiences to roll out and want help with a readiness review, a pilot sprint or governance, see our Databricks Genie implementation page.