Video · Bart Slodyczka ·
Anthropic Just Explained How to Build Always-On Agents (Master It In 10 Minutes)
In short
The video explains that always-on AI agents are usually dormant, waking for events, completing short runs, recording state, and sleeping again. It breaks down the five components required to build them and highlights how permissions, memory, and notification discipline make them useful and safe for background work.
Key points
- The five components are inbox, timers, memory, permissions, and outbox.
- Agents wake from events such as emails, user messages, payments, follow-ups, and schedules.
- Memory should store instructions, user preferences, prior messages, and task progress, using a compact index.
- Permissions should isolate credentials, restrict access, and require approval for high-impact actions.
- An invoice chaser can send approved reminders, ask for permission before contacting major clients, and schedule follow-ups.
- New events during a run can be queued, merged into the current task, or trigger a restart.
- Start with one task, such as a morning summary, and evaluate notification usefulness weekly.
Written by an AI model from the video’s captions and description.
Anthropic just published a guide on building always-on agents: AI agents that work for you in the background, sorting email, following up and messaging you only when it matters. I break down the five pieces every one of them needs, how they fit together, the mistakes to avoid, and how to build your first one.
In this story
A tool or model opens in the Heidelberg AI Curriculum: what it does, what it costs, how to start. A repo opens on GitHub.



