Building always-on agents
by Anthropic
The five primitives behind a stateful, proactive agent: event inbox, timers, memory, permissions and outbox.
Overview
Building always-on agents is a November 13, 2026 guide on claude.com written by William Hu of Anthropic, with thanks to Rodrigo Olivares and Michael Segner. It argues that an always-on agent runs the same core loop as any other agent and differs only in the architecture around it: the agent is idle most of the time and wakes for short runs, each of which receives events, does work inside a session transcript and saves state. The body is organised around five primitives. The event inbox is a durable table with an HTTP handler and a single background worker, with advice on fast acknowledgement, deduplication by source ID, batching pending events and handling events that arrive mid-run. Timers cover agent-set follow-ups and recurring schedules kept in a separate store and polled by a scheduler. Memory is a directory of files mapped to a database or object store, split into user instructions, user notes, past messages and per-task progress files, plus a 'dreaming' pass that curates memory after the fact, with prompt caching and compaction for the context window. Permissions use a credential-injecting proxy, Claude Code auto mode classifiers and deferred tool calls surfaced through MCP Apps. The outbox covers send tools, dispatchers and evals that tune message volume. There is no code; it is a design reference.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Guide
- Cost
- Free
- Duration
- ~14 min read
- Provider
- Anthropic
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Design a durable event inbox with a handler, a single worker and per-event status tracking
- ✓Choose whether mid-run events are queued, folded into the current run, or trigger a restart
- ✓Store follow-ups and recurring schedules separately from the inbox and poll them with a scheduler
- ✓Split agent memory into instructions, user notes, past messages and task progress files
- ✓Keep per-run context overhead small using memory indexes, prompt caching and compaction
- ✓Keep raw credentials away from the agent with a proxy that injects them into outgoing requests
- ✓Pause a run with a deferred tool call and resume the session when the user's approval arrives
- ✓Decide between a simple send tool and an outbox table with a dispatcher for agent messages
Highlights
- •Published November 13, 2026, and written against current Claude features such as auto mode, the memory tool and MCP Apps
- •Turns a vague product idea into five concrete primitives with storage and failure-handling advice for each
- •Covers operational details most agent tutorials skip: crash recovery, deduplication, one worker per inbox and message pacing
- •Vendor-neutral enough to apply with a durable execution engine or a custom scheduler on any stack
Who It’s For
Best For
- ✓Engineers building background or proactive assistants that act on webhooks and schedules
- ✓Platform teams designing memory and permission layers for unattended agents
- ✓Developers moving a chat agent to event-driven, long-lived operation
Prerequisites
- •Familiarity with the basic LLM agent loop of tool calls and results
- •Working knowledge of databases, webhooks and background workers
FAQ
What is Building always-on agents?
Building always-on agents is an Anthropic architecture guide for engineers who want an agent that keeps context across sessions and acts without a user watching. It breaks the design into five durable primitives, so afterwards you can build the event inbox, scheduler, file-based memory, permission layer and outbox that sit around an ordinary agent loop.
Is Building always-on agents free?
Building always-on agents is free to access.
What level is Building always-on agents for?
Building always-on agents is aimed at a intermediate audience. Recommended background: Familiarity with the basic LLM agent loop of tool calls and results, Working knowledge of databases, webhooks and background workers.
How long does Building always-on agents take?
Expect roughly ~14 min read. Most learners work through it at their own pace.
What will I learn from Building always-on agents?
You'll learn: Design a durable event inbox with a handler, a single worker and per-event status tracking; Choose whether mid-run events are queued, folded into the current run, or trigger a restart; Store follow-ups and recurring schedules separately from the inbox and poll them with a scheduler; Split agent memory into instructions, user notes, past messages and task progress files; Keep per-run context overhead small using memory indexes, prompt caching and compaction; Keep raw credentials away from the agent with a proxy that injects them into outgoing requests; Pause a run with a deferred tool call and resume the session when the user's approval arrives; Decide between a simple send tool and an outbox table with a dispatcher for agent messages.
Topics
Sources
This page was written from 2 sources, 1 on domains other than claude.com.