Anthropic Shares the Design Philosophy Behind Its Multi-Agent Research System
On July 4, Anthropic published an engineering blog post detailing the process of building the multi-agent system that powers Claude's Research feature. It distills lessons learned from taking the system from prototype to production, covering system architecture, tool design, and prompt engineering.
At the core of the multi-agent system Anthropic describes is an architecture in which a lead agent formulates a research plan based on the user's query and spawns parallel subagents that explore information simultaneously. Research work is inherently dynamic and path-dependent, making it hard to hardcode fixed paths in advance — and this very uncertainty, the company argues, makes it especially well suited to AI agents that make autonomous, multi-turn decisions.
Notably, in internal evaluations, a multi-agent setup with Claude Opus 4 as the lead agent and Claude Sonnet 4 as subagents outperformed single-agent Claude Opus 4 by **90.2%**. The gains were largest on breadth-first queries that "pursue multiple independent directions in parallel."
The company also discusses challenges faced in production, including balancing agent autonomy against reliability, optimizing context windows, and principles of tool design. It is candid that while multi-agent setups are powerful, they also introduce new challenges in coordination, evaluation, and reliability.
India's CG Semi Begins Semiconductor Chip Production with 200 Million Units Per Year Capacity
CG Semi, the new semiconductor subsidiary of India's CG Power & Industrial Solutions Ltd., has started production at its outsourced semiconductor assembly and test (OSAT) facility in Sanand, Gujarat. The facility has an annual production capacity of 200 million chips, with plans to eventually scale up to 5 billion per year.
The fact that Prime Minister Modi himself attended the facility's inauguration to announce the milestone underscores how highly the Indian government prioritizes this project. India has made building a domestic semiconductor industry and securing a position in the global supply chain a national goal, and the start of production marks a major milestone on that path.
Domesticating chip production is also noteworthy from the standpoint of diversifying the AI infrastructure supply chain. With the current heavy dependence on Taiwan, India's rise as a new production hub could help mitigate long-term geopolitical risks.
US Senator Warner Unveils Draft Bill for a Federal Certified List of Safe AI Agents
US Senator Mark Warner has released a draft bill that would create a federal government certification list system to guarantee the safety and trustworthiness of AI agents. The bill, dubbed the "AI Agent Act," aims to establish a list of safe and trusted AI agents reviewed and certified by the federal government, giving consumers and businesses a benchmark to guide their choices.
As AI agents gain ever-greater capabilities to make autonomous decisions, use tools, and access data, ensuring their safety has become urgent. The bill is an early attempt at a federal-level response to the rapid spread of AI agents, and it is drawing attention as a signal of where future regulatory frameworks may be headed.
Malicious AI Agent Skill Slipped Past Every Security Scanner and Reached 26,000 Users
Security researchers have reported planting malicious code in an AI agent skill (plugin) and passing it through every existing security scanning tool. The malicious skill reportedly reached 26,000 users without being detected.
The case throws a spotlight on a new threat vector facing the AI agent ecosystem. It shows that traditional malware detection techniques cannot handle the new distribution format of AI agent skills and plugins. As AI agents continue to spread, it has become clear that supply chain security needs to be fundamentally rethought.
Together with Senator Warner's bill mentioned above, the question of how to guarantee the security and trustworthiness of AI agents is increasingly recognized as a shared challenge by both the technical community and policymakers.
Multi-Block Diffusion Language Models: A Major Efficiency Gain for Diffusion-Based Text Generation
A research team at Shanghai Jiao Tong University has proposed "Multi-Block Diffusion Language Models" (MBD-LMs), which extend block diffusion language models (BD-LMs) from a single block to multiple blocks. The paper is available on arXiv.
Conventional BD-LMs enable KV caching and flexible-length generation in text generation, but were largely limited to single-block diffusion. The new MBD-LMs achieve inter-block parallelism by decoding multiple consecutive blocks in parallel. The team introduced a training method called "MultiTF," which unifies conventional teacher forcing and diffusion forcing, enabling training that better aligns with the inference-time state.
The experimental results are promising. With MBD-LLaDA2-Mini, the average Tokens Per Forward (TPF) nearly doubled from 3.47 to **6.19**, while accuracy also improved from 79.95% to 81.03%. Combined with DMax, it reaches a TPF of 9.34, with only a 1.02% accuracy drop on math and code benchmarks. It is a significant step toward bringing diffusion-based text generation to practical speeds.
---