--- title: "Google Rebuilds Colab Around Gemini to Run Full ML Pipelines Autonomously" subtitle: "Google's Colab gets a full AI-first overhaul powered by Gemini, turning the notebook into an autonomous data science agent with Learn Mode and Custom Instructions" url: "https://alphasignal.ai/news/google-rebuilds-colab-around-gemini-to-run-full-ml-pipelines-autonomously" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Google for Developers" company_url: "https://alphasignal.ai/company/googledevs" published: "2026-05-22T13:00:01.000Z" updated: "2026-06-16T21:09:08.060Z" type: "News" topics: - "Development" - "Llms" subtopics: - "Code Generation" - "Vision Language" upvotes: 306 read_time: "6 min" source_url: "https://developers.google.com/colab" comments: 0 --- # Google Rebuilds Colab Around Gemini to Run Full ML Pipelines Autonomously > Google's Colab gets a full AI-first overhaul powered by Gemini, turning the notebook into an autonomous data science agent with Learn Mode and Custom Instructions Google Colab has been a staple of the ML community for years: free GPUs, zero setup, shareable notebooks. But the tool you open today looks fundamentally different from the one you used a year ago. [The new Colab](https://developers.google.com/colab) is no longer just a hosted Jupyter environment. It's been rebuilt from the ground up as an AI-first coding partner, with Gemini embedded at every layer of the workflow. ## From notebook to agent The centerpiece of the new Colab is what Google calls the **Data Science Agent (DSA)**. This is not a chatbot that suggests the next line of code. It's an agentic system -- meaning it can plan, act, and reason across multiple steps without you holding its hand through each one. The DSA triggers autonomous analytical workflows: the agent creates a plan, executes code, reasons about the results, and presents its findings, all while allowing you to provide feedback and stay in control. In practice, this means you can drop a raw CSV into a notebook, type "clean this dataset, engineer features, and train a classifier," and watch Colab do it. From taking a raw dataset and asking the agent to autonomously perform cleaning and preparation, to generating feature analysis, training models, and evaluating the results -- this turns hours of work into a guided, conversational experience. ![Colab AI code transformation interface showing prompts and code blocks](https://developers.google.com/static/colab/images/effortless-code.png) The agentic loop the DSA runs through looks like this: 1. **Plan:** Outlines the steps it will take to accomplish your goal 2. **Execute:** Writes and runs the necessary Python code across multiple cells 3. **Reason:** Analyzes the output to inform its next steps 4. **Present:** Summarizes findings and surfaces them back to you 5. **Iterate:** Accepts your feedback mid-execution to refine or reroute the analysis ## The features under the hood Colab AI operates across your entire notebook, providing more accurate responses, taking actions across multiple cells, and engaging in meaningful ways to help you build faster. Beyond the DSA, the new Colab ships a set of tightly integrated AI features: - **Natural language code generation:** Generate functions, boilerplate, or whole-notebook refactors from a prompt - **Effortless code transformation:** Simply describe a change in natural language, and Colab will identify and refactor the relevant code for you. - **Inline error fixing:** When you hit an error, the AI doesn't just help find the bug -- it can suggest a fix in a clear diff view, helping you learn and keep going with your project. - **Autonomous visualization:** Data exploration is incomplete without visualization. Users are simply asking Colab AI to graph their data, and it generates high-quality, clearly labeled charts without the need for manual tweaking of plotting libraries. With previous Gemini integrations showing more than 2x efficiency gains in Colab, this AI-first reimagining can supercharge your workflows. ## Learn Mode and Custom Instructions Not everyone wants Gemini to just write the code for them. Google recognized this and shipped two features aimed at a different use case: learning and teaching. Custom Instructions, stored at the notebook level, allow notebook authors to tailor their Colab Gemini AI assistant and share their personalized assistants with the Colab community. Learn Mode turns Gemini into a personal coding tutor that provides step-by-step guidance instead of writing code for you. The key design decision here is portability. Custom Instructions are stored directly within your notebook and applied to future Gemini chats in that notebook. When you share your notebook with colleagues or the broader community, those instructions travel with it -- ensuring that anyone collaborating on or learning from your notebook gets the tailored AI experience you designed. Learn Mode is designed to provide "step-by-step guidance, instead of writing code for you," positioning Gemini as a tool for learning rather than only task completion. Learn Mode addresses concerns around over-reliance on generated code by encouraging users to engage with underlying concepts. For educators, this is a meaningful distinction: you can now ship a notebook where the AI is pre-configured to teach, not just answer. ![Colab workflow diagram showing connected steps in a notebook interface](https://developers.google.com/static/colab/images/interface.png) ## Who it's for and when to use it The new Colab is genuinely useful across a wide range of scenarios: - **Data scientists** who want to skip the boilerplate and go straight to insight -- let the DSA handle EDA, cleaning, and initial modeling - **ML researchers** who need to prototype fast without spinning up local environments or managing dependencies - **Educators** who can embed Custom Instructions and Learn Mode directly into shared notebooks, creating interactive coursework with a built-in AI tutor - **Students** learning Python or data science who benefit from step-by-step explanations rather than copy-paste answers - **Enterprise teams** who can use [Colab Enterprise](https://cloud.google.com/colab/docs/introduction), which adds IAM-secured workspaces, regional storage via Dataform, and full Google Cloud network integration ## The broader context: Google's agentic coding push Google I/O 2026 made one thing clear: AI coding is moving from autocomplete and chat inside the IDE toward autonomous development agents. Colab's transformation is one piece of a larger strategy. Google is positioning Gemini, AI Studio, Antigravity, Managed Agents, and Android integrations as a broader agentic development platform. The notebook format turns out to be a surprisingly natural fit for agentic AI. Cells are already discrete units of execution. The output of one cell is the input to the next. An agent that can read, write, and run cells is essentially operating in its native environment. By integrating the power of Gemini, Colab now functions as an agentic collaborator that can understand your code, intent, and goals, lowering the barrier to entry for tackling today's data problems. ## Pricing and availability The new Colab rolled out to the first users with more being added weekly. The core Colab experience -- including free GPU/TPU access and the base Gemini integration -- remains free. For heavier workloads, Colab Pro and Pro+ unlock more compute units per month, faster accelerators, more memory, and background execution up to 24 hours. On the consumer subscription side, Google AI Pro costs $19.99/month, while Google AI Ultra costs $99.99/month for power users who need maximum AI capabilities and agentic features. Enterprise teams can explore [Colab Enterprise](https://cloud.google.com/colab/docs/introduction) for IAM-secured, cloud-integrated deployments. You can start immediately at [colab.new](https://colab.new/). The real shift here is not just about convenience. Colab has always been where a huge chunk of ML education and prototyping happens. Embedding an agent that can autonomously run a full ML pipeline -- inside the same free, browser-based tool that students and researchers already use -- changes the baseline for what's expected from a coding environment. The notebook is no longer just a scratchpad. It's becoming a collaborator. ## Sources - [X post by @googledevs](https://x.com/googledevs/status/2057808700224446670) - [Meet the new Colab | Google for Developers](https://developers.google.com/colab) - [Supercharge your notebooks: The new AI-first Google Colab is now available to everyone](https://developers.googleblog.com/en/new-ai-first-google-colab-now-available-to-everyone/) - [Introducing Learn Mode: your personal coding tutor in Google Colab](https://blog.google/innovation-and-ai/technology/developers-tools/colab-updates/) - [AI-First Google Colab is All You Need - KDnuggets](https://www.kdnuggets.com/ai-first-google-colab-is-all-you-need) - [Google challenges Claude Code and OpenAI Codex with Antigravity 2.0](https://www.zleptnig.com/blog/google-io-agentic-coding-2026/)