Actualités IA Quotidiennes

PollyReach is a new tool that gives AI agents an actual phone number and voice, so they can make and receive real calls on your behalf. You set up the agent, define what it should say and do, and it handles the phone work — booking appointments, following up with leads, or fielding incoming inquiries.

The product is generating real buzz among automation builders today. The core appeal is simple: voice calls are still how a lot of business actually gets done, especially for service businesses. An AI that can hold a coherent phone conversation is genuinely more useful than one that can only send emails or fill out forms.

PollyReach supports integration with existing AI agent frameworks, so developers can plug in a phone capability without building it from scratch.

Cursor shipped Composer 2.5 today, which the company is calling its most capable model to date. Composer is the AI engine inside Cursor, the coding tool that has become one of the most widely used AI-assisted development environments among professional developers.

Composer 2.5 is focused on larger, multi-file code changes — the kind of refactoring and feature work where earlier AI coding tools often lost context or produced inconsistent results. Cursor says this version handles complex codebase-wide changes more reliably.

The launch is trending among developers today, arriving at a moment when the AI coding space is moving fast with new releases from multiple directions.

Railway, a popular cloud hosting platform used by thousands of developers, had its Google Cloud account suspended without warning yesterday. No prior notice, no grace period — just a sudden outage that took down customer apps across the board. Railway published a full incident report detailing exactly what happened and how they scrambled to recover.

The suspension appears to have been triggered automatically. Railway could not reach a human at Google Cloud in time to prevent the outage from spreading. They eventually got the account reinstated, but not before significant downtime had already hit their users.

The incident became a flashpoint in the developer community. Railway did nothing wrong — their account was simply caught by an automated system that didn't allow for human review before shutting off service.

Google announced a new pair of smart glasses at I/O yesterday — and the pitch is simple: they let you talk to Gemini without picking up your phone. The glasses have no screen. They listen, speak back, and connect to Google's full suite of apps and services, including Maps, Calendar, and Search.

Google is calling them "audio glasses" to signal that the experience is voice-first, not visual. You give a verbal command, and the glasses work through Gemini to get it done. The design takes a clear cue from Meta's Ray-Ban smart glasses, which have proven there is genuine consumer interest in wearable AI.

No price has been announced yet. Google said the glasses will be available to try at the event.

Google unveiled a new capability in AI Studio at I/O: describe what you want, and it generates a working Android app in minutes — no coding required. The tool runs in the browser and can produce native Android apps ready for testing and deployment.

This is one of the most direct challenges to traditional app development Google has announced. The target is clearly business owners, educators, and creators who have ideas but no engineering resources. You describe the app, the AI builds it, you ship it.

Google framed this as part of a broader push to make app development accessible to everyone — from teachers setting up classroom tools to small business owners building a customer-facing product.

Superpowers is an open-source framework for AI-assisted software development that has shot to the top of the Good AI List today. The project combines a skills library — a collection of reusable instructions for coding tasks — with a structured methodology for how developers should work alongside AI agents.

What makes it stand out is the emphasis on process, not just prompts. Most AI coding tools ask you to write better prompts. Superpowers asks you to change how you structure your development workflow — breaking work into discrete, reviewable steps that AI can execute reliably and humans can check.

The project has attracted a large and active community quickly, with developers sharing their own skills and refining the methodology through real-world use. It reads less like a toolkit and more like an opinionated answer to the question: how do you actually build software with AI agents?

Forge is an open-source reliability layer for running small AI models on your own hardware. Built by the AI Director at Texas Instruments, it addresses a specific and painful problem: when you ask a small AI model to use tools or call functions, it often fails in unpredictable ways. Forge adds a guardrail system on top that catches errors, validates outputs, and retries intelligently.

The numbers shared on Hacker News are striking. A standard 8-billion parameter model running agentic tasks hit 53% success rate without Forge. With Forge's guardrails layered on top, the same model reached 99%. That is not a small improvement — it is the difference between a tool that works and one that doesn't.

The project is domain-agnostic and tool-agnostic, meaning it can wrap around whatever model and whatever task you are running.

LobeHub launched a new release today positioning itself as the "Chief Agent Operator" for multi-agent workflows. The tool lets you coordinate multiple AI agents working on different parts of a task at the same time — one handling research, another drafting, another reviewing — through a single interface.

As AI agents get more capable, the challenge is less about what any single agent can do and more about how you coordinate a team of them. LobeHub is betting that coordination is the hard problem worth solving. The open-source project has a growing user base among developers building complex automation pipelines.

The latest release adds better visibility into what each agent is doing at any moment, which addresses one of the most common complaints about multi-agent systems: you lose track of what is happening.