Issue #165 12 stories

Global AI Weekly

OpenAI Pulls Models From Cursor

Published Tuesday, September 1, 2026

In this issue

Highlights

2 stories
OpenAI Pulls Models From Cursor
openai.com

OpenAI Pulls Models From Cursor

OpenAI says it will stop supplying models to Cursor following the AI coding company’s acquisition by SpaceX. The proposed cutoff is November 12, with OpenAI citing concerns that it cannot be confident SpaceX will comply with its contractual terms based on previous disputes involving Elon Musk’s companies. Cursor can continue using competing models, but the move highlights how ownership, platform competition and model access are becoming increasingly intertwined in the AI developer ecosystem.

Anthropic previews the Model Hardware Standard for AI-controlled devices
anthropic.com

Anthropic previews the Model Hardware Standard for AI-controlled devices

Anthropic has introduced the Model Hardware Standard (MHS), a model-agnostic specification that lets AI agents safely interact with physical equipment. Early deployments span microscopes, liquid handlers, robotic arms and even quantum-computing hardware. Rather than building bespoke integrations for every device, MHS provides a common interface that agents can access through protocols including MCP. Anthropic plans to develop safety evaluations with research and manufacturing partners before open-sourcing the standard.

In this issue

Research

2 stories
Skill Issue: Are Skills Language-Invariant in LLMs?
huggingface.co

Skill Issue: Are Skills Language-Invariant in LLMs?

The same AI model can become significantly better, or worse,. at reasoning simply by changing the language it uses. Researchers tested three models across eight languages and six games in more than 500,000 self-play matches. Performance varied systematically by language, even though the model, rules, and available actions were identical. Particularly interesting: switching only the model’s reasoning language recovered as much as 89% of lost performance in some experiments.

JIT-Agent: AI That Builds Its Own Agent Harness
huggingface.co

JIT-Agent: AI That Builds Its Own Agent Harness

Agent performance may depend as much on the surrounding harness as on the underlying model. JIT-Agent explores whether that harness memory, planning, tools, and execution strategy, can itself be generated and improved by AI. The system creates task-specific harnesses on demand and learns from previous executions. In the authors’ experiments, adding JIT-Agent produced substantial gains across several models, with DeepSeek-V4-Flash even outperforming GPT-5.6 on two agent benchmarks.

In this issue

Video

2 stories
Agents Don’t Know What They Don’t Know - Rob Zuber
youtube.com

Agents Don’t Know What They Don’t Know - Rob Zuber

The bottleneck in agentic development isn't the model. It's the feedback loop. When a developer writes code, the inner loop is tight: write, run, fix, repeat. When an agent writes code, that loop collapses and expands at the same time. Agents iterate fast. But without real feedback, they iterate confidently in the wrong direction. This talk traces the evolution of that problem and what CircleCI is building to address it. We'll walk through three primitives: stop hooks that pause agent execution at key decision points, review skills that bring automated judgment into the loop, and remote validation that moves CI infrastructure into the development environment instead of waiting for a push. We'll demo the progression hands-on: validate manually first, then let the agent take over that work automatically, then back it with remote compute. Each step is a forcing function toward the same goal: agents that fail fast, fix early, and ship code you can actually trust. The tools are already changing. The pipelines and workflows your team depends on are next.

The Model Isn’t the Agent: Why the Harness Matters
youtube.com

The Model Isn’t the Agent: Why the Harness Matters

IBM Technology explains why an AI model alone doesn’t determine how capable an agent really is. The surrounding agentic harness, tools, memory, orchestration, guardrails and execution loops can have just as much impact on real-world performance. The video clearly separates model intelligence from the infrastructure that turns that intelligence into action, and shows why two products using the same model can behave very differently

In this issue

Articles

2 stories
Agentic SQL for Free with Qwen3.8 27B and DuckDB | MotherDuck
motherduck.com

Agentic SQL for Free with Qwen3.8 27B and DuckDB | MotherDuck

If your laptop has 16GB of RAM, your agent can write SQL locally for free with Qwen3.8 27B. On the DABstep SQL benchmark, Qwen beat GPT 5.6 Luna and cost under 50 cents in electricity, over 17x less. Add DuckDB and you have a data analyst that runs on your own machine. This post covers the benchmark results and walks through the setup, including which quantized model to download for your laptop.

Agentic AI Needs Verification, Not Just Confidence
devblogs.microsoft.com

Agentic AI Needs Verification, Not Just Confidence

As agentic AI produces increasingly large volumes of convincing output, traditional “looks good to me” reviews no longer scale. Microsoft’s Julia Kordick proposes a practical verification framework combining deterministic tooling, AI-assisted checks, and human expertise. Using application modernization and AI-generated code as examples, she outlines layered validation approaches, from compilation, linting, and security scanning to domain-expert review, production traces, and shadow testing, to make AI output measurable rather than merely plausible.

In this issue

Upcoming Events

1 story
October 20–22: NVIDIA GTC Berlin
nvidia.com

October 20–22: NVIDIA GTC Berlin

Europe's biggest AI moment is happening this autumn. GTC Berlin brings together developers, researchers, and industry leaders to go deep on the full five-layer AI stack, from energy, chips, and infrastructure to open models and physical AI. The headline keynote by Jensen Huang takes place at the iconic Tempodrom on 21 October, with three days of sessions, hands-on labs, and networking at STATION-Berlin. For developers shipping agentic workflows or navigating open-source AI, GTC Berlin is where it all comes together. Eligible attendees get 25% off.

In this issue

Code

2 stories
How to Evaluate Live & Voice Agents in ADK
developers.googleblog.com

How to Evaluate Live & Voice Agents in ADK

Google has added native evaluation for live voice agents to its Agent Development Kit. This hands-on tutorial builds a multi-agent voice workflow in Python, creates simulated users, generates spoken interactions with Gemini TTS and scores complete conversations using natural-language rubrics. Evaluations can also run through the CLI or inside CI/CD pipelines, making it possible to detect regressions in agent behaviour rather than relying on manually testing voice demos.

In this issue

Podcast

1 story
Practical AI - Building the Foundation for the Agentic AI Era
open.spotify.com

Practical AI - Building the Foundation for the Agentic AI Era

Angie Jones, VP of the Agentic AI Foundation, joins Practical AI to discuss what needs to happen for agents, tools and systems to work together at scale. The conversation covers MCP, A2A, Goose, open standards and the need for neutral infrastructure around the emerging agent ecosystem. They also discuss enterprise adoption and where humans should draw the line between work they perform themselves and tasks delegated to AI agents.

Every Tuesday

Get the next issue in your inbox.

The AI links worth your time, curated by the community and always free.