Anthropic's guide to always-on AI agents

Anthropic broke down the architecture of always-on agents — event inbox, timers, memory, permissions, outbox — and showed how to build a simple one in Cowork without code on a Claude Pro or Max subscription.

Author: Michael Kokin ·

For an AI agent to work round the clock, the neural net itself doesn't need to think round the clock. Or burn through your token budget doing it.

On November 13, Anthropic published a guide on building *always-on agents* — autonomous agents that handle tasks on their own, watch for events, and ping a human when needed.

The idea is simple: the agent sleeps most of the time. It gets a signal, wakes up, does the task, remembers the result, and goes back to sleep.

What such an agent is made of

Anthropic identifies five core components:

How to try it yourself. No code, no n8n, no server of your own

Let's start by building a simple agent that puts together a daily AI news digest.

Step 1. Open Claude. In the Cowork section, find Scheduled → New task → Create with Claude. In the new Claude interface, you can create a task right from the chat.

Step 2. Paste in the instruction:

Every weekday at 9:00 AM Europe/Madrid time, check https://www.anthropic.com/news and https://openai.com/news/ for new posts from the last 24 hours.

Pick up to three of the most important AI tool news items. For each, note what happened, why it matters, the publication date, and a direct link to the primary source.

Prepare a short draft post for Telegram in Russian, up to 900 characters.

Don't make up facts. If there are no new publications, say so. If a source is unavailable, report it. Don't publish anything without my approval.

Step 3. Confirm the schedule with the Schedule button. Open the created task and hit Run now to check how the agent handles it. Tweak the instruction if needed.

Done. Claude will run the task in the cloud even when your computer is off — the same principle Anthropic built into Claude Code, so you can vibe-code with your laptop closed. Results will show up in the run history.

A paid Claude Pro or Max subscription is all you need for this. Tasks eat into your subscription limits, so you can't run them infinitely.

But what if you need a real autonomous bot?

Say you want the agent to not just check the news each morning, but react to new Telegram messages, remember conversations, schedule tasks on its own, and send results.

That's the next level. You'd need Telegram integration, persistent memory storage, and an event-triggered execution mechanism.

But even here, you don't have to write everything by hand. Claude Code can generate the code, tests, and config for cloud deployment. For your own app, you'll typically need hosting plus separately billed access to the Claude API.

Which, incidentally, is exactly what Anthropic's guide is about: how to build a system where the agent can work on its own not just once, but as many times as needed.

The core principle: it's not the neural net that works round the clock — it's the system that knows when to wake it up.

Scheduled Tasks instructions · Always-on agent architecture