Context

Notebooks run linearly, block by block. Deepnote's AI followed the same pattern. It could only insert new blocks forward from where you triggered it. It couldn't edit existing blocks, move things around, or touch anything above. But real notebook work isn't linear. Refactoring, documenting, fixing errors. These all require the AI to see and edit the full notebook, not just append new blocks to it.

Linear vs non-linear block execution

On top of that, Deepnote had three separate AI features, each added individually over time. A chat sidebar for asking questions about the notebook. Auto AI, positioned at the last block, for generating new blocks from where it was triggered. And Block AI, which let users edit a single block in place. This was an opportunity to bring them together into one interface.

Scope

Give the agent scope over the entire notebook, not just forward from the trigger point. And unify the fragmented AI features into a single interface.

Approach

We introduced edit mode. An agent with access to the full notebook that can edit, insert, and remove blocks. Before every action, the system takes a snapshot so you can always revert. After the agent runs, the chat lists what changed. On the notebook you see the diffs and can accept them individually or all at once. If you send another prompt, you accept the changes.

To unify the existing AI features, we moved editing capabilities from the notebook level into the chat. The interface splits into two modes: ask mode for questions, edit mode for the agent to take action.

Outcome

We shipped edit mode as a beta in late August 2025, keeping Block AI and Auto AI alongside it. GA followed in late September 2025. A large portion of users relied on existing AI features, so we moved them into the sidebar interface one by one rather than replacing everything at once. In future iterations, @mentioning specific blocks would retire Block AI, and pinning the agent to an insertion point would retire Auto AI.