MCP App StoreLe meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
The primary theme of the week was the transition from short-lived AI interactions to long-running, autonomous agency. We saw a coordinated push from the industry leaders to move AI from a "copilot" that suggests code to a "partner" that executes projects.
The most significant shift came from OpenAI with the launch of ChatGPT Work, an agent designed for long-running tasks across applications and files. This trend toward autonomy was echoed by Anthropic and Cognition, who demonstrated Claude Fable 5 running unattended for eight hours within Devin to deliver production-ready code. These aren’t just incremental updates; they represent a fundamental change in the AI operational model: from synchronous chat to asynchronous project management.
While the agents became more autonomous, the infrastructure became more "industrial." Google expanded Managed Agents in the Gemini API with remote MCP support and background tasks, addressing the stability requirements of production-grade agents. Meanwhile, Hugging Face focused on reducing deployment friction with one-click integrations for Amazon SageMaker and Microsoft Foundry, and pushing the boundaries of efficiency with Torch Attention Profile and native-speed vLLM backends.
The week also saw a pivot in model intelligence. Grok 4.5 signaled a shift from narrow software engineering toward broader, high-level intelligence. Simultaneously, the industry began a critical reckoning with how we measure success, with OpenAI questioning the signal-to-noise ratio in coding benchmarks like SWE-Bench Pro.
Key Highlights:
Hugging Face introduces a profiling tool for PyTorch attention mechanisms to help developers optimize memory and compute. This is crucial for building more efficient large-scale models.
.jpg)
Cognition integrated Claude Fable 5 into Devin, achieving a breakthrough where the model can run unattended for eight hours to deliver production-ready code. This marks a significant step in the reliability and autonomy of AI software engineers.
Google Cloud is expanding AlphaEvolve to help customers find the most efficient algorithms for complex problems like microchip design and logistics. It leverages AI to automate the discovery of high-performance algorithmic solutions.
GPT-5.6 is now the default engine for Microsoft 365 Copilot, enhancing capabilities across the productivity suite. This update brings higher quality output and faster performance to Word, Excel, and PowerPoint.
OpenAI has released GPT-5.6, offering higher intelligence per token and improved performance efficiency. This update aims to provide more on-demand capability for complex, high-ambition workloads.
OpenAI introduces ChatGPT Work, an agent capable of executing long-running tasks across various applications and files. This represents a shift toward more autonomous, project-based AI agents that can manage goals over several hours.
NVIDIA releases a new perspective and dataset focused on training data specifically for AI agents. A critical step for improving agentic reasoning and reliability.
OpenAI analyzes reliability issues in the SWE-Bench Pro coding benchmark. The findings highlight the need for more accurate evaluation methods for AI coding models.
Cursor introduces Grok 4.5, their most intelligent model to date. It represents a pivot from pure software engineering focus to broader intelligence capabilities.
Implementation of a native-speed vLLM backend for transformers. This aims to significantly reduce latency and increase throughput for large model deployments.
OpenAI has launched GPT-Live, a new generation of voice models designed for more natural, real-time human-AI interaction. The technology is now integrated into ChatGPT Voice, significantly reducing latency and improving conversational fluidity.
Hugging Face and Amazon integrate to allow one-click deployment of models into SageMaker Studio. This streamlines the pipeline from model discovery to production deployment for ML engineers.
Microsoft Foundry now supports Hugging Face models via Managed Compute. This integration provides developers with scaled, managed infrastructure to run open-source models within the Microsoft ecosystem.
Google introduces background tasks and remote MCP support for Managed Agents in the Gemini API. These updates enable developers to build more reliable, production-ready agents with better asynchronous capabilities.
Anthropic launches Claude Cowork for web and mobile, introducing an agentic workflow where Claude can perform tasks autonomously in the background. Users maintain control by reviewing and approving decisions.
Hugging Face released LeRobot v0.6.0, advancing the open-source robotics ecosystem. This update focuses on improved imagination, evaluation, and iterative improvement for robotic learning.
Hugging Face integrates with SkyPilot to offer zero-egress storage, allowing developers to run AI workloads across any cloud while keeping data on HF. This significantly reduces data transfer costs and simplifies multi-cloud deployments.
.jpg)
A guide on optimizing Claude Code by selecting the right model and effort levels. This helps developers balance speed and precision based on the complexity of the coding task.
Anthropic provides practical patterns for agentic coding using Claude Fable. The guide focuses on identifying 'unknowns' during the implementation lifecycle to improve agent reliability.
Hugging Face has revamped its Kernels system with major updates to improve performance and developer experience. These changes optimize the core compute paths for AI model execution.