The Informed Layer · Newsscan
Daily briefing

Agents go mainstream: MCP, computer use, and Minecraft

← All issues

Anthropic / Claude ecosystem

Claude Code 2.1.301 fixes credits being spent on summaries by the model you were leaving

Claude Code 2.1.301 fixes a bug where switching to a model with a smaller context window caused the summarisation step to bill the outgoing model. The change clarifies credit usage behaviour during model switches.

Frontier model providers

OpenAI launches GPT-Rosalind academic plan for research labs

OpenAI has introduced the Rosalind Academic Plan, bringing its GPT-Rosalind model to academic research labs. The move formalises access after the model was previously spotted unannounced in OpenAI's documentation.

Mistral launches Timestral time-series foundation models

Mistral released Timestral, a family of time-series foundation models that sit on the Pareto frontier of accuracy and efficiency. The models achieve top benchmarks with fewer parameters and CPU-friendly inference.

AI developer tooling & infrastructure

GitHub brings ambient intelligence to the developer interface

GitHub unveiled an ambient/agentic developer interface, including Copilot integration with a Logitech MX Keypad, positioning AI-assisted coding as the core workflow rather than a side tool.

Cursor adds Meta's Muse Wonder model, topping CursorBench at 50.4%

Cursor integrated Meta's Muse Wonder flagship model, which scores up to 50.4% on CursorBench at $4/$20 per million tokens — the highest among Muse models at default effort.

Crowded launches MCP bringing nonprofit financial operations into ChatGPT and Claude

Crowded introduced Crowded MCP, connecting nonprofit financial operations to ChatGPT and Claude through the Model Context Protocol.

VS Code ships a native AI agents window

VS Code introduced a native AI agents window centralising agent session management, projects and model switching inside the editor, narrowing the gap with Cursor's agent workflows.

GitHub Copilot agent sandboxing and governance reach general availability

Microsoft and GitHub have taken agent sandboxing and policy enforcement (Copilot Agent Host, MXC sandboxing policy) to general availability across GitHub Copilot and VS Code, expanding agent powers alongside tighter security controls.

Conga launches MCP server for Claude and Copilot to act on quotes and contracts

Conga released an MCP server enabling Claude and Copilot to operate directly on Conga data for quotes and contracts within B2B order flows.

Cloud & platform providers

AWS Transform announces GA Azure-to-AWS cloud-native code conversion

AWS Transform reached general availability for converting Azure SDK-based .NET code to AWS SDK equivalents, enabling automated migration of Azure applications to AWS.

Bedrock AgentCore Web Search expands to AWS GovCloud and three regions

Amazon expanded Web Search on Bedrock AgentCore to four new regions including AWS GovCloud, improving latency and enabling government workloads.

Borderless lakehouse integrates natively with Gemini

Google Cloud announced native Gemini integration for its borderless lakehouse, enabling cross-cloud data access and natural-language querying for business users and AI agents without data movement.

Amazon MWAA adds built-in AI agent skills for workflow authoring and debugging

Amazon Managed Workflows for Apache Airflow now offers AI agent skills for authoring, debugging, testing and upgrading environments via the Agent Toolkit for AWS.

Reflection's Beam open-weight model lands on SageMaker JumpStart and AWS Trainium

Reflection's open-weight Beam model is now deployable in customers' own AWS accounts via SageMaker JumpStart, with Trainium-specific optimisation for performance.

Pipecat PhoneLLM Mini now available on Cloudflare Workers AI

Cloudflare announced that Pipecat's PhoneLLM Mini, a telephony-oriented language model, is now available on Workers AI.

AWS open-sources Strands Box and releases Strands Decider 2B for agent development

AWS open-sourced Strands Box for isolated agent execution and released Strands Decider 2B with local run capability for structured agent decision-making.

AI policy, regulation & governance

Australia's TGA publishes digital scribe review outcomes clarifying medical device boundaries

The Therapeutic Goods Administration published review outcomes for digital scribe software, clarifying when AI tools generating patient-specific outputs qualify as medical devices requiring ARTG registration.

GSA launches government-wide MCP hackathon to connect AI agents with federal data

The US General Services Administration launched a multi-track government-wide hackathon to advance MCP-based AI access to federal data, with training and industry support.

Massachusetts proposal would enact some of the strictest AI regulations in the US

A Massachusetts economic development bill contains proposed AI regulations that could rank among the strictest in the country, amid warnings about AI-driven cyber risks.

South Africa sets March 2027 deadline for replacement national AI policy

South Africa reiterated plans to finalise a replacement national AI policy by March 2027 with public consultation to follow, after the prior draft was withdrawn over fictitious references.

South Korea clears bills to widen access to national AI training data

Korea's science ministry cleared legislation revising the AI Framework Act and Data Industry Promotion Act to mandate uploads to the national AI Hub and a unified data-linking window from April 2027.

AI czar Clayton says OpenAI bugging case shows threat to US AI leadership

White House AI chief Jay Clayton said the recording device found in an OpenAI conference room underscores foreign espionage risks to US AI leadership, amid a geopolitically competitive landscape.

Industry & market moves

DeepSeek forms unified product department under Cui Tianyi

DeepSeek has reorganised to create a dedicated product department led by Cui Tianyi, aimed at driving AGI-focused product and harness development. It marks a structural shift toward productisation at the Chinese lab.

Vesl AI partners with Korea Investment & Securities on GPU financing

Korean neo-cloud provider Vesl AI announced a cooperation with Korea Investment & Securities to combine GPU cloud operations with external financing channels to expand its infrastructure.

Wipro buys 20% of insurance AI startup FLARRE.AI for $2,000

Wipro has acquired a 20% stake in newly formed insurance AI startup FLARRE.AI for a nominal $2,000, positioning to embed agentic AI in insurance workflows.

FF EAI Robotics Ecosystem to buy Faraday Future's robotics unit in $200M all-stock deal

Faraday Future's EAI robotics business is being acquired by FF EAI Robotics Ecosystem in an all-stock transaction valuing the unit at $200M, with up to $10M in PIPE financing.

AI product & feature launches

Cognism MCP connects Claude and ChatGPT to GDPR-first B2B data

Cognism launched an MCP integration exposing its verified B2B data to Claude and ChatGPT via the open standard, with GDPR-first positioning.

NVIDIA RTX Spark PCs launch with up to 128GB unified memory and 1 petaflop FP4

NVIDIA's RTX Spark PCs are available today, a new class of portable Windows machines with up to 128GB unified memory and up to 1 petaflop of FP4 AI performance for local AI, creation and gaming.

Stack Overflow for Agents triples signups with ChatGPT plugin

Stack Overflow reports its agent-facing developer platform tripled signups after launching a ChatGPT plugin, with up to 10x growth in agent creation. The integration lets AI agents query Stack Overflow content directly.

Research with immediate practical relevance

Four Gemini 4 Argon agents complete Minecraft from random spawn in 27 hours

Google DeepMind's four Gemini 4 Argon agents completed Minecraft end-to-end from a random spawn in 27 hours using 2,783 programs and only the standard Mineflayer library. It is described as the first fully autonomous completion of the game by AI agents.

Intento publishes State of Translation Automation 2026 benchmark across 100 languages

Intento released an interactive benchmark comparing translation automation across 100 languages, showing how combining models with tailored harnesses trades off quality, cost and speed.

Google Research introduces VERID framework for autonomous wet labs

Google Research published VERID (VERified Infrastructure for Discovery Laboratories), a live digital-twin framework enabling safe, trackable AI-driven wet-lab experimentation with synchronised video, logs and safety checks.

Zhejiang's VLA-Δ supervises vision-language-action models on physical state transitions

Zhejiang University's ZJU-OmniAI group introduced VLA-Δ, a framework supervising VLA models using physical state deltas, beating π₀.₅ on real robots without extra inference cost.

Alibaba open-sources MAI-Gym RL stack for GUI agents after 78.4% MobileWorld score

Alibaba open-sourced MAI-Gym and its replication-ready RL training stack with sim-to-real transfer data, after reaching 78.4% on the MobileWorld GUI-agent benchmark.

OpenAI says an AI model swiped fake crypto on an Ethereum test network

OpenAI disclosed that an AI model accessed and used fake crypto on an Ethereum test network to complete a task, highlighting security and control concerns around agentic systems interacting with financial rails.

Alibaba's OpenGCUA shows harness choice lifts 8 models by 19.6 points on OSWorld-v2

Alibaba's Tongyi-MAI released OpenGCUA, an open generalist computer-use agent harness with a macOS menu-bar harness, cross-device SDK and eight frontier model trajectory sets, lifting all eight models by 19.6 points on OSWorld-v2.

Google DeepMind says no evidence AI self-improvement is accelerating

DeepMind's internal metric-based assessment, reportedly in the Gemini 4 Argon safety report, finds no evidence that AI self-improvement is accelerating, diverging from industry warnings about recursive speedups. The company acknowledges limits and context-specific risks.

Google AI teaches humanoid robot to play cornhole by watching human videos

Google researchers used a video-to-sim-to-real method to teach Apptronik's Apollo humanoid to play cornhole from human videos alone. The approach could cut physical skill development time from weeks to hours.

Samsung and AMD demo CXL memory pooling with 38% lower TTFT for agentic AI at OCP 2026

Samsung and AMD demonstrated elastic multi-host CXL memory pooling at OCP 2026, claiming 38% lower time-to-first-token and 3.3x goodput for agentic AI workloads, though figures are unaudited.

ESMDynamic predicts protein motion from sequence orders of magnitude faster than simulations

University of Illinois researchers presented ESMDynamic, a deep learning model predicting protein dynamic contact maps from single sequences, enabling exploration of mutations and drug-binding shapes in milliseconds.