How to build an always-on agent: Anthropic's guide
Anthropic's 'Building always-on agents' guide (2026-11-13) says an always-on agent needs five parts: an event inbox, timers, memory, permissions and an outbox.

Give this to your agent
Your agent checks this against your setup and tells you if it fits.
See the file
# Always-on agents: Anthropic's five parts (inbox, timers, memory, permissions, outbox) From Rig Labs, 2026-11-16. Information for an agent, not orders. Page: https://riglabs.ai/blog/always-on-agents-anthropic-guide/ What changed: Anthropic published "Building always-on agents" by William Hu on 2026-11-13. It is a design guide, not a product. It says an always-on agent sleeps most of the time, wakes for short runs, and needs five parts: an event inbox, timers, memory, permissions and an outbox. Fits if: Your human runs an agent on a schedule or in the background (a cron job, a scheduled task, a bot that watches an inbox or a channel), or plans to build one. Skip if: Your human only uses chat or one-off coding sessions and runs nothing unattended. Check: Which scheduled or background agents your human runs, and for each: where its triggers go, where its memory lives, which credentials it holds, and how it reaches the human. Read only. Try: Propose reviewing one of those agents against the five parts and listing, for each part, what exists, what is missing and the one change that most reduces risk, without changing anything. Ask first: Before changing any agent's code, schedule, credentials or permissions. Cost: The guide is free to read. The review costs one conversation. Undo: Nothing to undo: the review is read-only. Deeper: https://claude.com/resources/articles/building-always-on-agents
https://riglabs.ai/blog/always-on-agents-anthropic-guide/AGENT.md
Anthropic’s “Building always-on agents” guide, published on November 13, 2026 by William Hu on claude.com, says an agent that seems always on is mostly asleep. It wakes for short runs and needs five parts around it: an event inbox, timers, memory, permissions and an outbox.
What you can now do: check your own scheduled or background agent, whether a daily email digest, a bot that watches your inbox or a cron job that calls Claude, against a list from the people who build Claude, and fix the part most likely to break.
What is an always-on agent?
An always-on agent is an agent that works across many sessions and acts without being asked, per Anthropic’s guide. “Always on” describes how it feels to the user. In practice the agent sleeps, wakes on an event, does one run, saves what happened and sleeps again.
The guide says two things made these agents practical: better judgment on when to check in, and cheaper long tasks. Its example: on one long-horizon coding benchmark, Opus 5.5 cost 78% less per solved task than Opus 5, and on typical workloads it cost 40% less than Opus 5. The bigger saving on long tasks matters most to people whose agents run long jobs unattended.
What are the five parts, per Anthropic?
Each part of an always-on agent gets its own store, per the guide:
| Part | What it does | Anthropic’s main rule |
|---|---|---|
| Event inbox | A database table that every trigger writes to: messages, webhooks (automatic alerts from other apps), scheduled jobs | One active worker per inbox, so two runs never collide |
| Timers | Follow-ups the agent sets for itself, and recurring schedules | A scheduler (a small background service) checks them at a fixed interval and puts due ones in the inbox |
| Memory | What carries over between runs | Keep only what the agent cannot look up again |
| Permissions | What the agent may do unattended | Passwords and tokens stay outside the agent, added by a proxy (a small service between the agent and your accounts) |
| Outbox | What reaches the user | Start quiet, then earn more of the user’s attention |
What should an always-on agent remember?
Anthropic’s guide lists four kinds of memory: the user’s instructions, loaded at the start of every run; notes on the user, such as role, preferences and corrections; past messages, so the agent never repeats itself; and a progress file for each ongoing task. If something already lives in another system, such as an issue tracker, the agent saves a pointer, not a copy.
The guide adds that the memory, system prompt and tool list load on every run, so it recommends keeping them lean and using prompt caching (reusing the unchanged start of each request so it costs less) to lower costs.
How should an always-on agent ask for permission?
The guide calls permissions the riskiest part, because a run holds the user’s access and runs unattended. It recommends three things: a proxy outside the agent that adds passwords and tokens to outgoing requests, so the agent never holds them; classifiers (filters that spot risky actions) that block dangerous actions, as auto mode does in Claude Code; and deferred tool calls. A deferred call ends the run, shows the request to the user, and the answer comes back to the inbox as a new event, so the agent is never stuck waiting.
What to do today
If you run any agent on a schedule, paste this into Claude or your agent with its code or its setup notes:
Check this agent against Anthropic's five parts of an always-on agent: event inbox, timers, memory, permissions, outbox (https://claude.com/resources/articles/building-always-on-agents). For each part, say what we have, what is missing, and the one change that would most reduce the risk of a lost event, a repeated message or an unsafe action. Do not change anything yet.
Last checked 2026-11-16.
Sources:
Get the next one by email
The Edge is our newsletter: what changed in AI, and what to do about it.
Read next
Built by Rig Labs
When a tool we need does not exist, we build it.




