Claude Routines vient de sortir, et c'est DINGUE (Agents 24h/24 et 7j/7)

Anthropic has just released a new feature called "Routines" that transforms Claude into a full-fledged automation platform. This functionality allows users to define a prompt once, configure a workflow in three clicks, and have it run automatically on Anthropic's cloud environment without needing a computer to be on or dealing with tools like N8N or Make. Routines are intelligent automations that run autonomously. Users provide a highly specific prompt, choose when and how the routine is triggered, and Claude executes, reasons, and self-corrects if problems arise. Launched on April 14th, this feature is currently in a "research preview" or beta test phase. There are three ways to trigger a routine: 1. **Schedule:** Program the routine to run hourly, daily, weekly, etc. 2. **Webhook:** An external application triggers the routine. 3. **GitHub Event:** Changes or events in a GitHub repository trigger the routine. Multiple triggers can be combined for a single routine. An example use case is a developer working on an AI tool for a client; a routine could be triggered daily at 9 AM based on changes in their GitHub repo, such as a new feature or user, initiating an automated action. Routines differ significantly from existing market solutions. Unlike classic cron jobs, which simply execute a script at a fixed time, Routines are complete agents that reason using provided tools and can self-correct. Tools can be integrated via existing connectors on Claude (e.g., Gmail, Slack, Google Drive) or through standard APIs. A key distinction is that Anthropic hosts and executes everything. Previously, projects on Cloud Code or N8N automations ran locally or on an external, paid VPS. With Routines, Anthropic manages all hosting and execution, meaning users can close their PCs as everything is handled on Anthropic's servers, eliminating the need for a third-party VPS. This is included directly in the subscription. Subscription tiers offer different daily limits for activations and runs: * **Pro:** 5 activations, 5 runs per day. * **Max:** 15 activations, 15 runs per day. * **Team:** 25 activations, 25 runs per day. * **Enterprise:** 25 activations, 25 runs per day. Runs consume the subscription like a normal Claude Code session. There's a minimum interval: runs cannot be activated more frequently than once per hour. The core difference between N8N/Make and Anthropic's Routines lies in the "middle" processing. While both start with an event that activates a workflow and end with an output (e.g., email, Slack, CRM), N8N processes information linearly, node by node. With Routines, Anthropic intelligently manages everything autonomously, acting as an agent to achieve a defined objective. It's not a linear, fixed scenario; users explain what they want in natural language via a prompt, and Anthropic figures out how to achieve it. However, N8N remains a good choice for pure data transfer tasks that don't require reasoning, as it will be cheaper in terms of tokens. For simple, straightforward automations, N8N is recommended. If reflection and reasoning are needed, Routines become more suitable. Migration from N8N to Routines can be done by copying an N8N JSON configuration and asking Claude to transform it into a specific prompt for the routine. A live demonstration showed a routine designed to summarize emails and send them to Telegram. The user manually triggered the routine, which normally runs daily at 9 AM. The routine accessed the user's mailbox, classified emails into "promo" (marked as read) and "important" (left unread), and sent a summary of important emails to Telegram. This entire setup, from connecting to Telegram to configuring the AI assistant to read emails, was done in minutes. The routine's configuration includes: * **Prompt:** Defines the objective (e.g., "You are an assistant, read my emails, classify them into 'promo' and 'important', mark promos as read, leave important unread, and send a summary of important emails to Telegram"). Classification rules are also specified (e.g., emails with "unsubscribe," "offer," or from "no-reply" domains are promo). * **Environment:** Configured to allow network access. Anthropic offers different access levels (none, trusted, full, custom) to address security concerns about AI agents accessing entire systems. "Trusted" is for known connectors like Gmail, while "Full" is often needed for third-party APIs not directly supported by Anthropic's connectors. Environment variables, such as API keys and secret tokens (e.g., Telegram chat ID), are also set here. * **Model:** A model like Sonnet 4.6 can be selected for less complex tasks. * **Triggers:** Configured for scheduled runs (e.g., daily at 9 AM). * **Connectors:** Tools like Gmail are plugged in. Another routine was created from scratch for automated AI news monitoring. The objective was to retrieve the latest articles from RSS feeds of newsletters like Ben's Bites and TLDR AI, filter for articles published in the last 24 hours, synthesize the content, and send a structured summary to Telegram. The routine was set to run daily at 9 AM. During the creation of this second routine, a Telegram bot was set up to receive the summaries. This involved creating a new bot on Telegram, obtaining its token, and then getting the chat ID by sending a message to the bot and using a specific URL to retrieve update information. The bot token and chat ID were then added as environment variables in the routine's configuration, with the environment set to "Full" access for Telegram connectivity. When the news monitoring routine was executed, it attempted to access the RSS feeds. It successfully parsed TLDR AI but encountered issues with Ben's Bites due to a Cloudflare protection. Crucially, the AI agent demonstrated self-correction: it identified the problem with Ben's Bites and suggested finding alternative solutions, even proposing a service like RSS.app to generate a proxy RSS feed URL to bypass the issue. This highlights the agent's ability to reason and adapt without requiring manual intervention or complex N8N configurations. The routine eventually sent a detailed summary of AI news, covering tools, models, business, and trends, to the Telegram bot. While the initial output was long and untranslated, the user demonstrated how to easily modify the prompt to request shorter, more focused summaries or even translation, directly within the routine's configuration. The main takeaway is that Routines significantly lowers the barrier between having an automation idea and putting it into production. Users no longer need coding skills or the complexity of traditional automation software; they simply describe what they want, and the AI handles the execution, self-correction, and debugging. The critical question now is not "does it work?" but "what will you automate first?" Those who understand how to leverage this technology for business will gain a significant advantage.