GitHub has added project-level orchestration to the Copilot app. After users set a goal, Copilot can coordinate work across one or more code repositories and handle non-repository tasks. The system supports starting sessions from the project view, setting up recurring automations, and generating work artifacts, bringing previously scattered tasks together in one place.

Apple's research team proposed HALO, an autonomous control system for humanoid robots, and tested motions learned in simulation on the Unitree G1. The system uses a three-layer architecture: a vision-language model handles high-level task orchestration, the middle layer generates actions, and a motion tracker trained with reinforcement learning controls the robot's joints at the lowest level. Qwen3.6-35B-A3B is used by default for task orchestration.

World Labs has launched Frameboy, a cloud-based demo product that uses its autoregressive world model Atlas to generate scene visuals in real time as users move, streaming the results to a browser. Users can control the viewpoint and explore AI-generated 3D environments; cloud GPUs generate visuals at up to 15 frames per second.

Alibaba's AI coding platform Qoder has launched the Qoduck model, available to individual users for free for a limited time. Qoduck is a model option for the Qoder coding agent, aimed at users who develop with AI assistance.

Visa has upgraded its open-source security tool VVAH and introduced a related consulting service. Compatible with models from providers including Anthropic and OpenAI, the tool brings vulnerability discovery, analysis, remediation, and verification into a single workflow and provides feedback on scan progress. If an automatically generated patch fails verification, the system can continue refining it based on the feedback. The solution is said to shorten the vulnerability response cycle from weeks to hours.

Kling 4.0's Keyframes feature lets users provide up to 10 keyframes to control scene changes between the opening and closing frames, giving the video generation process more detailed storyboards to follow. Users can try the feature through Magnific.

Based on MiniMax-H3, HyperFlow H3 Flash uses architectural simplification, low-precision quantization, and a new sampling mechanism to reduce the model's size to as little as 9.1 GB. It can run on consumer PCs equipped with GPUs with large amounts of VRAM and supports real-time 768P video generation using four B200 GPUs.

Google's ReproBridge packages successfully reproduced machine learning experiments for reuse, including code, environment configurations, launch scripts, and documentation. In tests covering 36 papers, 1,029 tasks, and 4 models, the agent reproduction success rate rose from 76.4% to 92.6%, while API costs fell by 19.7%. Removing code or environment configurations caused the largest drops in success rate.

MaxxBench made minor changes to DeepSWE coding tasks, including changing the capitalization of filenames. Some models with original scores of around 74% showed score differences of up to 23 percentage points. In this test, Claude Opus 5.5 was the only model to maintain or improve its performance after the changes, highlighting differences in how well models adapt to changes in evaluation tasks.

SandBaseAI has raised several million dollars in seed funding to build personal agent infrastructure for developers. Its solution offers a unified API Key for accessing more than 2,000 models and APIs across over 100 platforms, while its Agent Runtime supports operations such as code execution and file handling. It aims to address the fragmentation of data, tools, and runtime environments in agent development.