Cursor Is Giving AI a Job After the Code Is Written

Morning, {{first name | folks}}! A lot happened in AI this week. Cursor added AI that works after deployment, Claude helped uncover a new enzyme system, Gemini gained more app connections, Fish Audio released a more controllable voice model, and Google added local AI support to Antigravity. Let’s get into it.

Cursor has launched two new bots for code after the pull request. Rollouts monitors changes from PR to production for regressions, while Security Reviewer checks every pull request for exploitable vulnerabilities. Both are available on Cursor’s Teams and Enterprise plans.

Rollouts can compare production signals with the state before a deployment and flag changes that cause problems with latency, errors, or other metrics. Security Reviewer looks across the codebase for issues, including injection, broken authentication, exposed credentials, and unsafe dependencies, then reports the attack path and suggests a fix.

Anthropic says Claude has discovered a previously uncharacterized enzyme system while searching through a massive database of DNA sequences. The system, called array-associated reverse transcriptases (ART), has a reverse transcriptase, a nearby partner gene and a long array of DNA repeats that resemble patterns found in CRISPR.

Claude used roughly 950 agents over 21 hours to examine more than 200,000 reverse transcriptases, narrowing them down to 20 candidates for detailed analysis. Anthropic’s scientists then tested the most promising candidate in the lab and are still working to understand what ART actually does.

Google is adding Airtable, Linear, monday.com, PandaDoc, and Zoho to Gemini, along with creative apps including Adobe, Picsart, Squarespace, and Webflow. The new connections let Gemini work with information and tasks from those services directly inside a chat.

Google is also adding lifestyle apps, including Apartments.com, Experian, Peloton, and SeatGeek. Users can connect the services through Gemini’s settings or call them in a conversation with an @ mention, expanding the number of tasks Gemini can handle without users switching between apps.

Fish Audio has released a preview of Drama 3, a new text-to-speech model that lets developers direct how an AI voice speaks using natural-language instructions. The model can change tone and emotion mid-sentence, switch between multiple characters, and regenerate individual words without recreating the entire recording.

Drama 3 is available through Fish Audio’s API as drama-3-preview, with access currently gated and pricing yet to be announced. Instead of treating text-to-speech as a single generation step, developers can now control different parts of a performance inside the same recording. That gives applications more control over dialogue, character delivery, and voice direction without stitching together separate audio clips.

Google has added local model support to its Antigravity SDK, letting developers run AI agents directly on their own machines. The SDK supports Gemma 4 26B A4B through LiteRT, along with local inference servers including Ollama, LM Studio, and vLLM.

That means developers can keep more agent workloads local instead of sending every request to a cloud model. Google is also showing hybrid setups where a cloud model handles the planning while a local model takes care of tasks such as code review, patching, and testing. For developers, the interesting question is starting to become which parts of an agent actually need a cloud model in the first place.