Cursor’s Agent SDK is a framework for building software agents tailored to a team’s workflows, rather than adapting to a fixed product or building all the infrastructure from scratch. Agents can respond to events across development work, such as pull requests, and apply policies teams can inspect and change. They can assess code risk, route reviews, approve changes that meet requirements, delegate tasks to specialized agents, or split work among parallel subagents. Built-in evaluations and event replay help teams test behavior and check that important workflows remain reliable.
Agents are structured like small software projects: instructions are written in Markdown, while tools and service integrations are defined in code, so changes can be version-controlled and reviewed alongside other software. Cursor reports that one internal team built a pull-request reviewer in about a day; it checks changes against a knowledge base of past lessons, cites relevant notes, and flags contradictions. The SDK is used by PGA America for its software development factory. Teams can develop and test agents in a local playground, then deploy them to Cursor’s infrastructure with one command, managing secrets, integrations, logs, and evaluations through a dashboard. Agents can also run on a team’s own infrastructure as containerized Node processes. The SDK is generally available and includes guidance for coding agents to assist with setup.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Google AI Mode sometimes avoids directly saying Joe Biden won the 2020 US presidential election, instead directing users to official sources. Its responses vary across users and election years, while answers become more direct when asked outside the US. (arstechnica.com)
AI data centers can consume millions of gallons of water daily, competing with local communities for drinking water as their numbers grow. Residents around the world are organizing to oppose projects that threaten local water supplies. (restofworld.org)
Cartesia argues that tokenizers limit end-to-end learning because they impose hand-crafted text boundaries. Its hierarchical networks learn to group raw bytes into useful units, offering a scalable path toward byte-level language models. (cartesia.ai)
Particle-1 is a software-engineering model post-trained on DeepSeek V4.1 Flash that uses about 45% fewer tokens while maintaining comparable accuracy. Across four coding and data-engineering benchmarks, it cuts average task costs by 34.4%-51.6%, with plans to expand into other domains. (particle.ai)
Google’s Gemini gave different results for two election-related queries that differed by one digit, according to an X post. The post claims one result censored Joe Biden but does not show the queries or responses. (twitter.com)
Google introduced Gemini 4 Carbon, a new model in its Gemini lineup. The announcement describes the model and its capabilities. (blog.google)
Today's best Hacker News stories, summarized and screenshotted, one email a day.