Google’s new developer AI tools

Google is showing a broader push to turn its models into working developer tools, from autonomous coding agents to local model setups and document-heavy research workspaces. (blog.google ) (ai.google.dev) (notebooklm.google) The coding piece builds on Jules, Google Labs’ asynchronous coding agent, which Google introduced in December 2024 and opened to everyone in August 2025. Google says Jules reads a repository, plans multi-step changes, and proposes code updates developers can review before merging. (blog.google 1) (blog.google 2) Reports circulating this month describe “Jitro” as a more proactive follow-on to Jules, with a demo showing an agent that can evolve code toward broader goals instead of waiting for one prompt at a time. Google has not published a standalone official product page for Jitro, but recent coverage and social posts frame it as the next step in the Jules line. (devops.com) (testingcatalog.com) (x.com) That shift follows Google’s own description of how developers are already using scheduled “pods” of Jules agents for recurring jobs like performance tuning, security patching, accessibility fixes, and test coverage work. In a December 2025 post, Google said the Stitch team had made Jules one of the largest contributors to its repository through those background tasks. (blog.google) A second piece of the push is local use of Gemma, Google DeepMind’s open-weight model family. Google’s current documentation shows developers can run Gemma through Ollama on a local machine, and Google says that path can be useful for experimental or low-volume applications, including a locally run code assistant. (ai.google.dev 1) (ai.google.dev 2) Google updated the Gemma line again this month with Gemma 4. Google’s release pages say the new family includes E2B, E4B, 26B A4B, and 31B variants, supports text and image input, and offers context windows up to 256,000 tokens. (ai.google.dev 1) (ai.google.dev 2) The research side of the story is Google’s notebook-style workflow for grounding answers in user-provided material. NotebookLM says it is built around sources users choose, and Google’s help pages say the system answers based on uploaded material rather than pulling unsupported claims from outside the notebook. (notebooklm.google) (support.google.com) Posts about “Gemini Notebooks” describe a merger between Gemini and NotebookLM-style source management, including support for more than 100 files in a project workspace. Google’s consumer NotebookLM plans page separately says the product can synthesize hundreds of sources, while older public limits for standard notebooks were lower, indicating Google is expanding how much material these tools can hold. (x.com) (notebooklm.google) (medium.com) Google is also adding more document-handling plumbing under the hood. Its Gemini API documentation says developers can process PDFs up to 1,000 pages and store uploaded files for 48 hours through the Files Application Programming Interface, which points to the same larger push toward long-context, source-backed work. (ai.google.dev) (ai.google.dev) Taken together, the pieces point in one direction: Google wants developers to choose between cloud agents like Jules, local Gemma setups through Ollama, and source-grounded notebooks for research-heavy work, without leaving Google’s model stack. The open question is how much of the more autonomous “Jitro” behavior becomes a public product rather than a demo. (blog.google) (ai.google.dev) (x.com)