Claude Code team at Anthropic has released a new context engineering framework: for Claude Opus 5 and Fable 5 models, the team removed over 80% of system prompts, with coding evaluations showing no measurable performance degradation. Key principles include replacing rigid rules with judgment-based guidance (e.g., changing 「do not write comments」 to 「write code consistent with surrounding style」), using tool interface design instead of example stacking, progressive disclosure with on-demand loading of Skills, and replacing hand-written CLAUDE.md with automatic memory. The team argues that as model autonomy improves, overly detailed rules can encroach on decision space, introduce conflicts, and increase inference cost. Some developers have already applied these principles to reduce personal configurations from 23,500 tokens to 9,259—a ~60% reduction.

XinZhi Embodied and Fudan University jointly released the N0 series tactile technology report and the NeoData dataset, containing over 30,000 hours of vision-tactile interaction data, approximately 1.4 million operation segments, 3.3 billion timesteps, 8 billion RGB frames, and 10 billion tactile frames, covering six robot embodiments and 450 real-world long-horizon tasks, with 5,000 hours open-sourced. The NeoForce unified representation model resolves incompatibility issues across different tactile sensors. N0-VTLA achieves an 85% success rate in plug-and-play tasks (vision-only: 60%) and 99% in key-pulling tasks (vision-only: 35%) by predicting the next 50 steps of tactile evolution. N0-TWAM integrates tactile sensing into a world model using a 7.2-billion-parameter Mixture-of-Experts architecture, achieving an average simulation success rate of 84.5% (baseline: 36%) and an average real-robot task success rate of 46.3% (LingBot-VA: 21.9%).

Kuaishou's KwaiKAT team has released the agentic coding model KAT-Coder-V2.5, trained across more than 100,000 verifiable code repository environments, leveraging a novel AutoBuilder and data expansion flywheel. The model treats agentic coding as an infrastructure challenge rather than a matter of model scale, ensuring correctness through execution within task descriptions, executable environments, and verification tests. The team found that about 16% of training failures stemmed from sandbox infrastructure issues rather than model strategy, highlighting the importance of isolated execution environments. To prevent overfitting and test-case exploitation, they employ process-level prompting and randomized scaffolding (e.g., tool names, formats), along with asymmetric PPO and multi-layer reward shaping for complex behaviors—where the Critic can access privileged context such as test coverage.

Nanbeige Lab has released Nanbeige4.2-3B, a 3B-parameter Looped Transformer model that achieves unified capabilities in coding, office tasks, complex tool usage, and general reasoning through architectural innovation. After a standard forward pass, the model feeds hidden states back into the same Transformer layers for a second computation; this two-loop design optimally balances performance, speed, and stability, delivering approximately 75% improvement in token efficiency compared to standard Transformers of the same size. Training involved building an agent data pipeline covering real-code-repo sandboxes, live online MCP services, and cross-file dependency office datasets, using a multi-stage post-training strategy. The model consistently outperforms larger models such as Qwen3.5-9B and Gemma4-12B in evaluations across General Agent, Code Agent, reasoning, and alignment benchmarks.

Tencent Cloud has launched CodeBuddy NPC, a development-process-native intelligent agent that uses Git repositories as native memory and is deeply integrated with the CNB platform, advancing from AI-assisted coding toward autonomous closed-loop delivery. By integrating Git, CI/CD, and other stages, the agent enables an automated 'commit-validate-repair-revalidate' loop. Through trimming redundancy, standardizing SOPs, and optimizing tool calls, initial token consumption was reduced from over 20,000 to around 2,000—a drop exceeding 90%. The platform supports multi-role NPC collaboration, enabling enterprises to build virtual teams comprising product managers, technical leads, developers, and testers for parallel workflows, transforming developers from 「code producers」 to 「supervisors who define goals, make decisions, and validate outcomes」.

A static analysis study of 25 AI Agent frameworks identified 30 paths where model-controlled parameters can directly reach high-risk operations—such as file write/delete, shell execution, network egress, SQL execution, and message sending—without any authorization checks, revealing a critical security gap in 「validation ≠ authorization」. Most agent tools only verify whether request formats are correct, but fail to check at runtime whether the current caller is authorized to perform the specific action. CrewAI had the most findings (12), followed by SuperAGI (6) and MetaGPT (5). The initial scan achieved an 81% precision rate, with 12 false positives fixed and regression tests added. The authors emphasize that identified paths do not equate to confirmed vulnerabilities, nor prove exploitability—only that no authorization checks exist along these paths. Custom agent loops were not included in the scope.

Alibaba's Tongyi Qwen AI platform has launched, offering a unified API to access over 100 models (integrating Qwen, Kimi, DeepSeek, etc.). New users receive free credits, and developers can invoke diverse models using a single API Key. The platform open-sourced two official Skills (qianwen-ai for model selection, invocation, and billing; qianwenai-deploy for application deployment), enabling agents to autonomously select models and deploy applications via natural language commands. In testing, an agent produced a 15-second short video—from text storyboard to image generation, video synthesis, speech synthesis, and audio-video merging—with a single input, invoking models like wan2.6-t2i, wan2.6-i2v-flash, and qwen3-tts-flash, at a total cost of approximately ¥5.35. The platform emphasizes agent-centric design, dynamically switching model sizes based on task complexity to balance cost and performance.

Simon Willison summarized an investigation into an underground LLM API token resale market, where model access is resold at discounts by abusing free trials, stolen credentials, and open-source proxy software. China's relay market leverages open-source tools like one-api and new-api to load-balance illegally obtained credentials and resell them at reduced prices. Buyers seek lower costs, circumvent regional restrictions, or collect data for model distillation. The investigation highlights the abuse risks facing publicly exposed LLM endpoints—once discovered, entire ecosystems profit, leading to massive token bills. The author urges LLM providers to implement strict hard spending caps per API Key, automatically halting service when a dollar threshold is reached, protecting developers from abuse-driven cost explosions.