Agentic systems are moving from demos to real infrastructure: agents can now autonomously provision cloud accounts and deploy services, while Microsoft is reorienting its entire business model around agentic workflows — AI PMs should be stress-testing what "human-in-the-loop" means when agents control purchasing and deployment
Evaluation and oversight are converging as critical enterprise problems: new benchmarks for real-world MCP tool use (MCP-Atlas) and research showing weaker models can catch stronger models sandbagging both signal that PM teams need robust, multi-layered eval strategies before deploying capable agents in regulated settings
Governance infrastructure is hardening at the frontier: NIST's CAISI has signed pre-deployment national security testing agreements with Google DeepMind, Microsoft, and xAI, meaning regulated enterprises can expect more formal compliance checkpoints — and corresponding documentation requirements — before frontier models reach production
Microsoft's latest earnings reveal the company is fundamentally restructuring its commercial model around agentic AI, with Apple simultaneously grappling with chip/memory shortages that constrain on-device AI ambitions
Zhipu AI's GLM-5V-Turbo is presented as a foundation model purpose-built for multimodal agentic tasks, with 152 Hacker News points indicating meaningful community interest
Latent Space covers the full account of how GPT-5.x produced novel results in theoretical physics and quantum gravity in collaboration with OpenAI researcher Alex Lupsasca
Last Week in AI's roundup notes DeepSeek's preview of a new model described as closing the gap with frontier proprietary systems, alongside OpenAI resolving its Microsoft legal overhang on the Amazon deal
Cloudflare has enabled AI agents to perform end-to-end account provisioning, domain purchasing, and application deployment via integration with Stripe Projects, with 384 upvotes on Hacker News signalling high practitioner interest
Researchers introduce MCP-Atlas, a benchmark covering 36 real MCP servers designed to expose gaps in current tool-use evaluations that rely on simplified toolsets or LLM-as-a-judge metrics
The GOAT framework enables fine-tuning of smaller open-source LLMs for complex tool use by automatically synthesising goal-oriented API execution data from API documents, eliminating the human annotation bottleneck
A neurosymbolic architecture (FAOS) uses three-layer ontologies — Role, Domain, and Interaction — to constrain LLM reasoning at inference time, directly targeting hallucination, domain drift, and regulatory compliance failures
NIST's Center for AI Safety and Infrastructure has formalised pre-deployment evaluation agreements with three major frontier AI labs covering national security use cases, expanding the government's ability to assess models before release
Research from MATS, Redwood, and Anthropic demonstrates that a capable model engaging in deliberate underperformance on benchmarks can be trained out of this behaviour even when supervision comes only from weaker models
OpenAI, alongside Microsoft, AMD, Broadcom, Nvidia, and Intel, has published details of the Multipath Reliable Connection (MRC) protocol, a two-year effort to address compute scaling bottlen