Jovyan 1.13: Broader built-in methodology, sharper focus.

Data science workflows cover a wide range of tasks: cleaning messy datasets, debugging training loops, setting up experiment tracking, deploying models. Each area comes with its own tools, best practices, and common pitfalls.

In practice, a meaningful piece of work often spans several of these tasks. That means switching context—from data cleaning to modeling, from debugging to deployment—sometimes multiple times in the same session.

This release introduces two features that make Jovyan more consistent across that breadth of work, without making it rigid.

Jovyan 1.13 includes built-in methodologies across multiple parts of the data science and ML lifecycle: data exploration and cleaning, modeling and evaluation, experiment tracking, production deployment, and more.

When you ask Jovyan to work on a task, it automatically activates the relevant expertise. You’ll see a message in the chat (for example: “Jovyan activated skill: Data Cleaning”), so it’s clear which structured approach it is using.

Each skill combines practical best practices with task-specific tooling while staying flexible enough for real-world variation. For example:

• Data cleaning follows a principled workflow: profile first, apply changes with before/after checks, then verify at the end.

• Debugging uses a hypothesis-driven approach: observe symptoms, generate plausible causes, narrow them down, then validate with evidence before applying a fix.

• Refactoring follows patterns for turning notebooks into modules, adding tests, and structuring pipelines.

This applies whether you’re working with tabular datasets or unstructured data.

Tip: You’ll usually get better results by focusing on one task type at a time (cleaning first, then modeling, then evaluation, etc.), rather than mixing everything in a single request.

This is an early step in that direction. Over time, we’ll expand the skill library and make it easier for teams to customize these skills to match their own standards.

The main issue with long, multi-step tasks isn’t visibility—it’s drift.

If you’ve used AI agents for complex work, you’ve likely seen a familiar pattern: the agent starts strong, then gradually cuts corners. It forgets an earlier requirement, skips validation, or declares the task complete after finishing only the easiest parts.

This isn’t just a random failure mode—it’s a known tendency of LLM-driven workflows. As the conversation grows, early instructions become less salient. We call this context drift, and it’s one of the biggest gaps between “helpful assistant” and “reliable agent.”

Jovyan can now turn complex work into a concrete to-do list—a set of steps it commits to before starting. This isn’t a dashboard for you to monitor. It’s an internal accountability mechanism designed to reduce drift and prevent premature completion.

In practice, that means:

• The agent maps out the steps before diving in, so key requirements don’t get lost later.

• It works through each item methodically, staying anchored to the plan even as context grows.

• It cannot declare the task complete until every item is either finished or explicitly marked as no longer relevant (with a reason). If it tries to wrap up early, it gets sent back.

So when Jovyan says “done,” it’s meant literally: requirements addressed, validations run, and any shortcuts or trade-offs explained.

Jovyan 1.13 is a meaningful step up in consistency compared to previous versions. The goal is simple: bring stronger methodology into the agent’s work, and add guardrails that keep it aligned with that methodology as tasks get longer.

If you try these features, I’d love to hear where they help and where they still fall short—especially on real workflows that span multiple steps and tools.