Cursor Launches Projects Feature: A Single AI Agent Manages Thousands of AI Agents for Large-Scale Development

The "Third Era" of software development is accelerating. The renowned AI code editor Cursor officially launched its new feature - Projects (beta version) on September 10th. This breakthrough feature aims to help developers handle large and complex tasks such as functional development, large-scale migrations, and complete application building. By introducing a "coordinator agent" to coordinate thousands of sub-agents to run in parallel, it frees developers from the tedious management of agents, allowing them to directly command the tasks themselves.

In terms of core architecture and operational mechanism, Projects introduces three core capabilities. First, default cloud execution with on-demand local testing: each Project runs on an independent cloud computer, and even when closing the laptop, it will not be interrupted, thus supporting massive parallelism of sub-agents beyond local computing power. When local testing is required, the coordinator automatically starts the local agent. Second, shared system context: each Project maintains a set of continuously synchronized files, where agents will deposit their research results, codebase understanding, and preferences. Over time, this makes the coordinator increasingly efficient. Finally, an intelligent subscription mechanism: the coordinator can monitor Slack channels, run on a schedule, or track all Pull Requests (PRs), taking proactive actions when signals are detected without requiring continuous manual prompts.

Regarding specific application scenarios, Projects mainly covers three practical models. In functional development, the system first has agents conduct research and establish a shared context, after which the coordinator formulates a plan and assigns sub-agents to implement and test in parallel. After going live, it can also continue monitoring logs and handling bugs. In code migration, for large-scale framework upgrades or style replacements that are easy to start but hard to finish, the system can gradually proceed through hundreds of PRs. Initially, human review is strict, while later, as the solution matures, the coordinator autonomously continues the process. For ongoing "gardening" maintenance, it is particularly suitable for handling code quality monitoring or regression testing, such as automatically extracting design system components and generating rules online to achieve efficient automation governance.

Cursor's official team has been conducting in-depth testing of Projects for several months and has applied it to migrations and feature releases involving hundreds of PRs. Test data shows that using this feature significantly increases productivity: the number of PRs merged by new users increased by 30%, while for users primarily relying on Projects, the number of PRs merged reached an astonishing six times. Currently, this feature is in beta stage and will be gradually rolled out to all users starting today.

Xiaomi Open Sources Industrial-Level Target Speaker Speech Recognition Large Model CocktailASR-1

Xiaomi open sources the industrial-level target speaker speech recognition large model CocktailASR-1, aiming to solve the cocktail party problem where multiple people speak simultaneously. The model changes the traditional ASR input method, accurately locking onto and recognizing the target speaker's voice in noisy environments, improving recognition capabilities in complex scenarios, and providing new solutions for the industry.

Bidirectional Communication and Smooth Interaction: OpenAI Launches the Full-Duplex Speech Model GPT-Live-1

OpenAI has launched the full-duplex speech model GPT-Live-1, which uses a single-model architecture to replace the traditional 'speech-to-text - reasoning - text-to-speech' cascade process, significantly reducing latency and enabling human-like smooth interaction. The model is now available via API for developers, and it can handle complex dialogue scenarios such as pauses, interruptions, and topic shifts more robustly.

Samsung and Mistral Collaborate to Empower Chip Design and Manufacturing

During the South Korea-France Presidential Summit, Samsung Electronics announced a collaboration with French AI startup Mistral AI, integrating its cutting-edge services and solutions into its semiconductor business to empower chip design and manufacturing. Mistral AI was founded in 2023 by former core employees of Google's DeepMind and Meta, and has high-performance large language models such as Mistral Large.

Runway ARR Breaks $200 Million: The Video Trend Subsides, Company Bets on World Models

AI startup Runway's annual recurring revenue breaks $200 million, disclosed on September 8th; four months ago, it added over $40 million in net new revenue in a single quarter, the fastest growth in history. Its milestone significance is not in the video business itself, but in its strategic shift: from selling text-to-video tools to building a simulation layer.

Anthropic Invests 51.7 Billion Dollars in 11 Months for 14.8 GW of Remarkable Computing Power

The arms race for AI computing power is intensifying. Anthropic signed a computing power contract worth 51.7 billion dollars in 11 months, adding at least 14.8 gigawatts to its existing 1 to 2 gigawatts starting from October 2025, and launching its own data centers. However, its scale may still fall short of OpenAI, reflecting the strategic confrontation behind its aggressive expansion.