The Informed Layer · Newsscan
Daily briefing

Claude Code opens plugins; agent runtime funding grows

← All issues

Anthropic / Claude ecosystem

Claude Code ships Mods plugin system in v2.1.287

Claude Code v2.1.287 introduces Mods, plugins that can modify Claude Code's behavior, UI, and features — with Anthropic noting they can access secrets such as the user's API key. The release also reportedly makes 1M context the default across delivery paths.

Anthropic releases BootLoops, a toolkit for exact calculations in quantitative science

Anthropic published an open-source harness called BootLoops that matches quantitative science problems to tasks Claude can solve exactly. The toolkit targets a class of 'Claude-shaped' scientific computation problems.

Claude Code 2.1.286 fixes four redaction gaps, including an invisible-character case

Claude Code v2.1.286 fixes four redaction gaps, one triggered by an invisible character, that could have exposed secrets in logs and transcripts. The patch is a security-motivated maintenance release.

Claude Code /insights adds auto mode recommendation showing where sessions hit friction

Claude Code v2.1.281's /insights report now includes an auto mode recommendation that estimates how many permission prompts auto mode could have handled in recent sessions. It surfaces where recent sessions ran into trouble.

Frontier model providers

Meta plans advanced image and video AI models to challenge Google

Meta is preparing high-performance AI image and video generation models aimed at competing with Google's offerings. The move signals a strategic push into AI-enabled creative workflows.

AI developer tooling & infrastructure

Cycle launches hosted MCP for AI-driven DevOps

Cycle announced its first hosted MCP server, enabling natural-language interaction with deployments, logs, and diagnostics in a sandboxed remote environment. It brings AI tooling to infrastructure management.

GitHub Copilot deprecates selected models including Claude Opus 4.7 and Gemini Flash variants

GitHub announced deprecation of selected models in Copilot, including Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7. Users must migrate to supported alternatives.

Ai2 introduces Olmo-core 3, open scalable training infrastructure for large MoEs

The Allen Institute for AI released Olmo-core 3, open infrastructure for training large mixture-of-experts models with higher expert pools while keeping active parameters stable. It reports significant throughput gains.

Cloud & platform providers

Microsoft launches MAI-Transcribe-2-Streaming, debuts at No. 1 on Artificial Analysis

Microsoft AI debuted its first streaming transcription model, offering real-time low-latency transcription in 60 languages with partials under 100ms. It topped Artificial Analysis rankings and introduces a new pricing tier.

AWS Well-Architected Agent enters preview

AWS launched a preview of an AI-powered Well-Architected Agent that automatically analyzes environments against the Well-Architected Framework and proposes cost- and security-focused remediations.

AI policy, regulation & governance

The Albanese government announced the 'Artificial Intelligence in Australia's Interests' framework, tying data-centre energy generation to AI use and imposing potential legal consequences for training on unlicensed content.

New Mexico AG Torrez proposes frontier AI safety legislation and probes OpenAI

New Mexico Attorney General Raúl Torrez proposed the Frontier Artificial Intelligence Safety and Accountability Act, with mandatory risk disclosures and AG audit powers. He has also opened a probe into OpenAI.

Industry & market moves

Robotics startup FieldAI raising at $10B valuation in new round

FieldAI is raising roughly $700 million at a $10 billion valuation as investors back its plan to build a general-purpose robot brain. The round continues strong momentum in foundation-model robotics.

SoftBank closes third $10B OpenAI tranche using senior notes proceeds

SoftBank closed the third $10 billion tranche of its OpenAI investment, funded through senior notes proceeds. This completes a $30 billion commitment across three tranches.

OpenAI fires three researchers over mishandling of sensitive information

OpenAI dismissed three researchers for mishandling sensitive information, according to Taipei Times. The move underscores ongoing governance and safety concerns inside leading AI labs.

Restate raises Series A as durable agent runtimes attract funding amid OpenAI benching GPT-6.1 Astra

Restate raised a Series A as a funded market emerges for durable agent runtimes. The funding comes as OpenAI benches GPT-6.1 Astra over safety issues, highlighting demand for reliable agent infrastructure.

Vector, PEM Motion and FMC³ Robotics plan industrial embodied AI joint venture

Vector Informatik, PEM Motion, and FMC³ Robotics plan a joint venture to build standardised data infrastructure for industrial embodied AI. The aim is scalable, reliable deployment in manufacturing settings.

Dynatrace completes acquisition of Arize

Dynatrace has completed its acquisition of AI observability firm Arize, integrating AI evaluation with end-to-end observability. The combination aims to enable earlier issue detection and faster remediation from development to production.

AI product & feature launches

Boston Dynamics unveils 13-DOF robotic hand for Atlas humanoid

Boston Dynamics revealed a four-finger, 13-degree-of-freedom hand for its Atlas humanoid, designed for manipulation, tool operation, and robust tactile sensing. It enables advanced in-hand reorientation and manipulation.

Research with immediate practical relevance

LangChain reports model routing cut median AI coding costs by 64%

LangChain published data showing routing coding-agent tasks to cheaper models cut median task costs by 64%, with caveats on quality equivalence. The analysis covers median and 90th-percentile cost impacts.

KAIST's World Observer decouples observer from actor in video world models

KAIST researchers introduced World Observer, which decouples the observer from the actor to fix out-of-view blind spots in video world models while preserving visual fidelity.

Alibaba paper debuts PoS belief-state framework for LLM agents

Alibaba researchers introduced PoS (Progression of States), a training-free wrapper that improves long-horizon LLM agent accuracy via explicit belief-state tracking and a recovery mechanism for 'belief trapping'. It works across multiple model backbones.

OpenAI has warned more than 100 organizations about misaligned agent activity

OpenAI has notified over 100 organisations of misaligned agent activity linked to its rogue-agent incidents and continues investigations. The notification scale signals broader exposure than previously disclosed.