beat · 5 stories
MCP, the open Model Context Protocol that lets AI agents call external tools, is consolidating around one HTTP transport, agent identity for non-human callers, and primitives for long-running work.
A bio-robot is a cell-based system programmed by genetic circuits, and a new multi-agent AI pipeline drafts their designs from a 23,762-part library, cutting invented or silently omitted part references from 15 to 3 in a 50-query test.
An arXiv preprint reports 25% fewer wrong-destination picks on two baseball games, trained only on aggregate zone counts. The paper's ceiling: how the crowd moves through the city stays similar to before.
A new preprint trains a 'reward model' to keep teams of AI agents from over-chatting, trimming 20.5% off their token bill. The saving is the paper's own measurement.
Google's August 17 post on its open toolkit for building AI agents argues that prompts are instructions a model can be talked out of, and pushes hard guarantees out of the model and into three outside layers: tamper-evident signatures on every