Agentic AI is moving fast from code to knowledge work and physical services: coding agents are gaining persistent goal-tracking, and vertical AI agents (call handling, ER triage) are hitting unicorn valuations and clinical benchmarks that regulators will scrutinize
Healthcare AI is simultaneously showing strong diagnostic performance and attracting new safety architecture research — a signal that the bar for production deployment in regulated verticals is rising on both the capability and governance fronts
Both OpenAI and Anthropic are actively restricting access to sensitive AI capabilities (cybersecurity, inference chips), reinforcing that supply-side constraints and safety-motivated gatekeeping will continue to shape what enterprise AI PMs can actually ship
OpenAI, having publicly criticised Anthropic for limiting its Mythos model, has now imposed its own access restrictions on the Cyber model, reversing its earlier posture
A study using electronic records and brief nurse notes found o1 outperformed human triage physicians on emergency diagnosis, with researchers calling it a "profound change" for medicine
Anthropic is reportedly exploring Fractile's custom inference chips as a supply hedge alongside Google, Amazon, and Nvidia, driven by explosive sales growth straining existing infrastructure
OpenAI's Codex CLI now lets developers set a /goal that the agent pursues across turns, using injected prompt files to evaluate completion or halt when a token budget is exhausted
A practical guide details how to wire the Google Workspace MCP server into Claude Code and Codex, enabling agents to read and act on enterprise productivity data
A breakdown of the architectural pattern that converts a base LLM into an autonomous agent, covering planning, memory, tool use, and execution loops
The startup's agents manage inbound calls and dispatch for HVAC, plumbing, and similar trades, reaching unicorn status across three funding rounds
An open-source CLI tool enabling AI agents to control native desktop applications directly, surfacing at 91 points on Hacker News
Researchers propose a context-aware multi-agent system that detects conditionally correct but medically inappropriate LLM responses, targeting hallucination and unsafe agreeableness in healthcare
This paper argues the field is undergoing a paradigm shift from narrow RL agents to open-ended autonomous agents built on LLMs, and examines what that means for training and reward design
A new multi-agent architecture separates "depth" reasoning (single target) from "breadth" aggregation (many entities), enabling schema-consistent outputs at web scale