Spark AI goes agentic: retrieve, reason, and act on your workspace

Until now, Spark AI answered questions. From today, it plans them: agentic mode breaks a request into steps, searches across your notes, tasks, and transcripts, and shows its reasoning as it goes.

Ask “what did we promise the client before the summer break?” and Spark AI will list the meetings involved, read the relevant transcripts, cross-check the tasks that came out of them — and answer with sources. Every step is visible, every claim is linked.

How it works

When a question needs more than one step, agentic mode takes over:

1. Plan — the request is broken into concrete search steps.
2. Retrieve — notes, tasks, and transcripts are searched in order.
3. Reason — findings are cross-checked and contradictions flagged.
4. Answer — with every source linked, every step visible.

You see the plan unfold in real time. “Listing meetings before July 1 — 4 found. Reading transcript Client review (06/26). Cross-checking 6 linked tasks. Done — 3 commitments found.” No spinner, no black box — a reasoning trace you can open, inspect, and challenge.

“If you can’t audit the reasoning, you can’t trust the answer. Agentic mode makes the thinking part of the product.”

That principle shaped every design decision. If a step returns nothing, you see it return nothing. If two sources disagree, the contradiction is flagged instead of silently resolved. The answer is only as good as the trail behind it, so we made the trail the interface.

Only your data, still

Agentic mode doesn’t change our rule: Spark AI reasons exclusively on your workspace — layers 01 and 02 — on Mistral, hosted in Europe. No public web, nothing shared across organizations.

• Answers drawn from layers 01 + 02 only — never the open web
• Every chat scoped to its project’s permissions
• Inline code stays intact: GET /api/v2/chats returns the full reasoning trace

Permissions deserve a special mention. An agent that can search everything is only safe if it respects who can see what. Every retrieval step in agentic mode runs under the permissions of the person asking. If you can’t open a document, Spark AI can’t read it for you — not in step one, not in step four.

What to try first

The best way to feel the difference is to ask something you’d normally reconstruct by hand:

• “What decisions did we take about pricing this quarter, and who owns the follow-ups?”
• “Summarize everything we’ve told this client about the migration timeline.”
• “Which tasks from last month’s retro are still open, and where did they come from?”

Each of these needs multiple sources, and each answer arrives with its receipts: the meetings, the notes, the tasks, linked inline.

Agentic mode is rolling out to all plans this week — no configuration needed. It’s available today in every Spark AI chat: just ask something that needs more than one step, and watch it think.