Multi-agent.

An orchestrator routes your request to specialists — Data, Analyst, Viz, Report — each focused on one slice of the work. They share kernel state, so variables travel from one to the next.

Switching modes

Toggle from the agent panel header — there's a small mode switch next to the model picker. The current conversation carries over; new turns route through the orchestrator.

Want to switch back? Same toggle. The thread keeps both runs and you can see which mode each turn ran under from the label next to the speaker name.

The four specialists

Each specialist has a narrower toolset and a sharper focus:

• Data — schema exploration, source discovery, SQL drafting, joins, profiling.
• Analyst — Python analysis, statistics, transformations, hypothesis checking.
• Viz — chart selection, encoding, formatting, annotations.
• Report — narrative prose, summaries, executive write-ups, markdown structure.

You don't pick the specialist — the orchestrator does. If you have an opinion ("use the Viz agent here"), the orchestrator honors it. Most of the time you just ask the question and the routing happens behind the scenes.

You          Compare Q1 revenue to Q4 by channel and write up the result.

Orchestrator → Data        pull the per-channel quarterly revenue
Data         [Tool: query]  ... 24 rows
             → handoff to Analyst

Orchestrator → Analyst     compute % change, flag channels with > 10% shift
Analyst      [Tool: execute_code]  ... created df_changes
             → handoff to Viz

Orchestrator → Viz         chart the deltas
Viz          → Proposed: 1 chart cell  [Review]
             → handoff to Report

Orchestrator → Report      write a 3-paragraph readout
Report       → Proposed: 1 markdown cell  [Review]

Shared kernel state

Specialists work against the same Python kernel as your notebook. When Data runs a SQL cell and assigns the result to orders, Analyst sees orders in scope without re-querying. Viz can grab the DataFrame Analyst computed. Report can read any variable to ground the prose.

That shared state is what makes multi-agent useful — handoffs are nearly free because no one re-fetches what someone else already has.

Approvals in multi-agent

Same model as single agent. Read-only tool calls run without prompting; writes pause for explicit approval. The approval banner shows which specialist requested the write — useful when you want to know which step is about to touch the database.

See Approvals for the full safety story.

When single agent is the better call

Multi-agent is heavier than single. For quick questions, single agent wins on latency. For exploratory work where you don't yet know which specialist applies, single agent is also simpler. Switch to multi-agent when:

• The task spans clear phases (pull → analyze → chart → write).
• You want a tidy log with named roles for a teammate to review.
• You'll iterate on individual phases without restarting the whole flow.

Back to Agents overview, or read about tools and approvals.