China Enacts the World's First AI Agent Regulations — Defective Agents Can Be "Recalled"
On July 15, 2026, the world's first dedicated framework regulating AI agents by name came into force. What drew attention is that it was neither the EU nor the US — China moved first.
A defining feature of the new framework is that defective AI agents can be subject to "recalls," much like automobiles. It is seen as a design that clearly delineates "decisions only humans can make" and applies graduated restrictions on agent autonomy.
Only limited details about day-one operations have been made public, and specifics on real-world effectiveness and enforcement mechanisms await future directives. That said, it is significant that the regulations target "agents" rather than "models" — this is arguably the first serious attempt to define the legal status of systems that autonomously choose and use tools. As discussions around agent-specific risks advance in Japan as well, this looks set to become an important precedent worth comparative study.
Top 7 Claude Code Skill Searches Are All About "Restraint"
According to an analysis that gained traction on Qiita, the top 7 Claude Code skill search trends were all "restraint"-oriented rather than about "adding capabilities." In other words, demand for making agents "not do unnecessary things" now exceeds demand for adding new features.
The author recounts an experience of "handing in a one-line bug report and getting back a 400-line pull request." A new abstraction layer, two helper modules, and configuration options nobody asked for were generated one after another, while the actual fix — all four lines of it — was buried somewhere in the middle. This is probably the "over-implementation problem" many developers have experienced firsthand.
The theme also connects to the Chinese regulations discussed above. While legal regulation moves in the direction of "stopping runaway agents," practitioners on the ground are seeking skill designs that "keep agents from doing too much." The bottleneck for practical AI agent deployment seems to be shifting away from raw capability toward control and narrowing the scope of action.
xHC: Expanded Hyper-Connections Extends the Scaling Axis of the Residual Stream
Drawing attention on LocalLLaMA is a new architecture proposal called "xHC (Expanded Hyper-Connections)." Its distinguishing feature is raising the Hyper-Connections method — which expands the Transformer's residual stream into N parallel streams — practically to N=16, beyond the conventional ceiling of N=4.
The research team identifies two bottlenecks as the reason existing HC variants (such as mHC) hit a wall at N=4: one is insufficient write-back information, and the other is residual mixing costs that grow cubically with N. xHC solves both challenges by enriching write-backs through temporal feature expansion and adopting a sparse design that updates only k=4 of the N=16 streams.
In evaluations on 18B and 28B MoE models, the 18B MoE improved average downstream task scores by 4.0 points over mHC. In scaling law experiments, vanilla and mHC reportedly require 1.50x and 1.19x the compute of xHC, respectively, to reach the same loss. Furthermore, by combining it with "xHC-Flash," which reduces memory traffic, the benefits of large-scale residual expansion can be enjoyed at a memory cost comparable to N=4.
The impression is that residual stream expansion is becoming a realistic candidate as the next scaling axis for LLMs, following "width, depth, and data volume."
Comparing 5 AI Coding Subscriptions on Price, Limits, and Use Cases
Canopy Wave has published an article comparing five major AI coding subscriptions by pricing model and usage limits. It is also generating buzz on Hacker News.
The services covered appear to be Claude Code, Cursor, GitHub Copilot, Cody, and Windsurf (due to limitations of the article body, details are not included in the collected data). Points of interest include differences between pay-as-you-go and flat-rate pricing per plan, context length limits, daily request caps, and differences in enterprise-oriented features.
The article emphasizes that it is not a matter of simply "picking the cheapest" — the optimal choice changes depending on your working style (whether you focus on long-form reviews or on generating fine-grained diffs). For organizations considering adoption, differences in security and data-handling terms will also become important.
CPO: A New Policy Optimization Framework Attuned to Answer Quality
"Contrastive Policy Optimization (CPO)," featured on Hugging Face Daily Papers, is research that introduces a new perspective into advantage design in reinforcement learning.
The core of the proposal is to use the distributional discrepancy between "reference-guided generation" and "vanilla generation" as a token-level correctness signal. The research team provides theoretical backing for this intuition, and further shows that on-policy self-distillation can be interpreted as a special case of CPO.
The practical significance is that it may enable evaluating generation accuracy at a finer granularity in settings where reward signals are noisy or partial. The benefits are expected to be especially large for tasks with clear ground-truth answers, such as mathematical reasoning and code generation.
AI Watermarks Are "Not Yet Ready" as Courtroom Evidence
"AI Watermark Evidence Fails Forensic Readiness," an empirical study published on arXiv, systematically evaluated whether watermarks embedded in AI-generated text can hold up as evidence in legal and forensic contexts.
The conclusion is, to put it mildly, harsh: current watermarking technology has not reached the level usable as courtroom evidence (forensic readiness) in terms of tamper resistance and detection reliability. Relying solely on current watermarks as the basis for platform operators claiming "this text is AI-generated" is deemed risky.
This also raises an interesting point in connection with the aforementioned Chinese agent regulations. How to position "proof of AI generation" within legal frameworks is a difficult challenge intertwined with freedom of expression and Due Process, and there appears to be a considerable gap between the state of the technology and institutional demands.
---