Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Anthropic Releases Guide: A Five-Component AI Agent Runs 24/7 Without Supervision

Anthropic has published a guide to AI agents that run 24/7, meaning around the clock without a human at the screen. Such an agent wakes on a schedule or an event, such as a new email, completes its task, and goes back to sleep. It remembers where it stopped, survives restarts, and asks permission before risky steps. The guide outlines five key components, a solid foundation for moving from demo to production.

AI-processed from TG: @neuraldvig (Нейродвиж); edited by Hamidun News
Anthropic Releases Guide: A Five-Component AI Agent Runs 24/7 Without Supervision
Source: TG: @neuraldvig (Нейродвиж). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @neuraldvig (Нейродвиж) ↗
Editor-in-Chief
Key metric #1Wakes on a schedule or event and sleeps between tasks
Saves progressafter a failure, it resumes from where it stopped
Key metric #3Dangerous actions, such as payments or deletions, go through a request to a human

Most AI agents today work like a conversation partner: you give a task, the agent works while the tab or terminal stays open, and that's it. But the most useful automation looks different: sorting through email in the morning, checking a failed deployment overnight, reconciling reports on Fridays. For that, an agent has to run 24/7, meaning around the clock and all seven days a week, without a human nearby. Anthropic has released a guide to building such always-on agents. It covers five key components, the architecture, and implementation details. For the full list of components, it is best to read the original source.

What an agent that runs 24/7 is

Running 24/7 does not mean the agent sits in memory and burns tokens nonstop. It wakes up when the time comes or an event happens: an email arrives, a test fails, a new task appears in the tracker. Then it does the work and goes dormant again. In everyday terms, it is not a lamp someone forgot to switch off but a motion sensor: most of the time it does nothing and costs nothing, but it reacts on time.

According to Anthropic's description, such an agent has three important properties. First, it remembers where it left off, so the next run does not start from zero. Second, it can continue work after a restart: a server crash or an update does not erase progress. Third, before risky actions, it asks a human for permission.

Why it is harder than it looks

Anyone can schedule a script with cron. The difficulties start when a language model making decisions stands behind the script. To leave an agent unsupervised, several engineering problems need solving.

The first problem is state memory. If an agent loses its connection midway through a long job, it must know what has already been done and what has not. Otherwise, after a restart, it will send the same email again or reapply a database migration. That is why progress should be stored outside the model itself: in a file, a database, or a queue.

The second problem is trigger-based launching. A schedule and an event are different mechanisms, and it helps the agent to understand why it was woken up so it does not have to guess at the context.

The third problem is the limits of autonomy. An agent working at night cannot ask a colleague for advice, so permission for dangerous actions becomes part of the architecture. Deleting data, making payments, and publishing on the company's behalf should logically go through human confirmation, while the rest can run on autopilot.

What practitioners should do with this

Three takeaways are useful for developers and product managers.

First: design the agent as a process with persistent state, not as one long dialogue. If a task cannot be safely interrupted and resumed, it is too early to release it into production.

Second: draw up in advance a list of actions that require approval. A good rule is to divide operations into reversible and irreversible ones and ask a human only about the latter. Otherwise the agent will flood you with requests, and the point of autonomy disappears.

Third: start with narrow scheduled tasks, such as a daily report or link checking. Move on to reacting to events later.

Limitations

Anthropic's material is an architecture guide, not independent research. The summary contains no figures on reliability or savings, so it cannot be used to judge how such agents perform in real conditions. The guide is likely aimed at the Claude ecosystem, so some recommendations will need adapting if you use other models or your own orchestrators.

Still, the fact that one of the leading model developers is publishing instructions for agents that run without constant supervision points to where the industry is heading. Attention is shifting from how smart the model is on a single request to how reliably the infrastructure around it is built: memory, triggers, recovery, and risk control. For teams planning to automate routine work, it is a useful starting point.

Story

Key Questions Answered

What an agent that runs 24/7 is?

Running 24/7 does not mean the agent sits in memory and burns tokens nonstop. It wakes up when the time comes or an event happens: an email arrives, a test fails, a new task appears in the tracker. Then it does the work and goes dormant again. In everyday terms, it is not a lamp someone forgot to switch off but a motion sensor: most of the time it does nothing and costs nothing, but it reacts on time. According to Anthropic's description, such an agent has three important properties.

Why it is harder than it looks?

Anyone can schedule a script with cron. The difficulties start when a language model making decisions stands behind the script. To leave an agent unsupervised, several engineering problems need solving. The first problem is state memory. If an agent loses its connection midway through a long job, it must know what has already been done and what has not. Otherwise, after a restart, it will send the same email again or reapply a database migration.

What practitioners should do with this?

Three takeaways are useful for developers and product managers. First: design the agent as a process with persistent state, not as one long dialogue. If a task cannot be safely interrupted and resumed, it is too early to release it into production. Second: draw up in advance a list of actions that require approval. A good rule is to divide operations into reversible and irreversible ones and ask a human only about the latter. Otherwise the agent will flood you with requests, and the point of autonomy disappears.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…