Jovyan 1.10 - A dedicated Data Cleaning agent, smarter agent handoffs, and smoother daily workflows

We are excited to announce Jovyan 1.10. This release introduces a new Data Cleaning agent : a specialized mode designed for data quality work; and sets the foundation for more specialized data science agents in the future.

If you’ve ever lost trust in an analysis because “something felt off” in the data, you already know the real problem: most of the time, the bottleneck isn’t modeling. It’s getting to a dataset you can confidently use, explain, and share.

Data cleaning is where real-world projects slow down. Not because it’s hard to write a few transformations, but because it’s hard to be sure you’re doing the right transformations, and that they didn’t quietly introduce new issues.

A good cleaning process answers questions like: Are these missing values expected? Are those duplicates real duplicates or legitimate repeated events? Are categories consistent? Are dates parsed correctly? Are keys safe to join on? And if we change anything, can we prove the dataset is now healthier?

Jovyan 1.10 gives this work its own dedicated agent so you don’t have to force a general-purpose coding workflow into a data quality mindset.

We built the Data Cleaning agent around a simple idea: data quality work needs a repeatable methodology, not a bag of tricks.

So instead of jumping straight into “fixing,” the agent starts by exploring the dataset in a structured way: it looks for patterns in missing data, suspicious distributions, inconsistent formats, and category fragmentation (the classic “NY” vs “New York” vs “New-York”). It tries to understand what the dataset is supposed to represent, and where reality diverges.

From there, it proposes cleaning actions that are deliberately conservative: the goal is not to make the dataset “pretty,” but to make it reliable. Changes are applied in a way that stays readable in notebooks and easy to review, and the agent keeps track of what it changed and why.

Then comes the part that matters most: verification. After each meaningful change, the agent checks that the dataset improved in ways you can observe: fewer invalid values, consistent parsing, stable keys, and sanity checks that prevent accidental distortions. When a decision depends on business meaning that the data alone can’t reveal, the agent flags it explicitly instead of guessing.

Finally, the agent documents the result: what was wrong, what was fixed, what remains uncertain, and what you might want to monitor next time. The outcome is a dataset you can use with confidence and a trail of reasoning your team can trust. This also create the deep understanding and context for later works, whether it is analyses or modeling, both for collaborators and AI agents.

There are two ways to use Data Cleaning in Jovyan 1.10.

1) Use it directly when your intent is clearly data quality: “clean this dataset,” “validate this extract,” “standardize dates,” “deduplicate customers,” or “find inconsistencies before modeling.” In this case, Data Cleaning mode will guide the work end-to-end and deliver both cleaned outputs and a clear cleaning report.

2) Start in Code mode, and branch into Data Cleaning when needed. In practice, data cleaning is often a branch of coding: you begin a notebook or script for analysis, and quickly realize the data needs attention. From Jovyan 1.10, Code mode can now suggest switching to Data Cleaning when it detects that the core task is data quality. You keep the momentum, switch to the right mindset, and then return to Code mode once you’re ready to move on to modeling, feature engineering, or deployment.

This makes the workflow feel natural: the right agent takes over at the right time, without breaking your project flow.

Want to see a demo? Check out the match between AI agents for data cleaning here.

Jovyan 1.10 also improves how agents collaborate. When you start from a high-level design with Architect mode request, Jovyan can now more reliably route execution to the right next step. Either a straightforward build path with Code mode or a more structured multi-step approach with Orchestrator mode so you get less back-and-forth and more steady progress from plan to working result.

Alongside the headline features, Jovyan 1.10 includes several reliability improvements that remove common friction points:

More reliable notebook execution. If an agent creates a new notebook as part of your workflow, it now opens correctly in the editor so running cells works smoothly. No more “it exists but doesn’t run” moments.

Fewer annoying edge-case errors when agents read notebooks. In previous versions, certain unusual notebook states could trigger an error when the agent tried to inspect cells. In 1.10, the agent handles these cases gracefully so the workflow keeps going instead of breaking unexpectedly.

Cleaner handling of escaped characters in tool calls. Sometimes models (typically Gemini 3) output slightly inconsistent escaped characters. Jovyan 1.10 is more robust to these variations, which means fewer “weird parsing” hiccups when tools are used in complex tasks.

Better access to config files. Small JSON/YAML config files now show up properly for the agent instead of being treated like opaque data blobs. That means easier setup, fewer “I can’t see the config,” and smoother project initialization.

Data Cleaning mode is the first step toward a broader direction: specialized agents for specialized data science work, each built around a clear methodology and designed to hand off smoothly to other modes.

If you spend meaningful time debugging and cleaning datasets, this release will definitely make the process much faster, saving your time and creative energy for higher valued works.