Anthropic details five building blocks for always-on agents · 2026-11-14
Archive backfill: evidence edition 2026-11-14, added on 2026-11-14. These are not today’s announcements.
(November 13 backfill) Five primitives support persistent agents and resumable approvals
On November 13, 2026, Anthropic published “Building always-on agents,” describing five primitives around the agent loop: an event inbox, timers, memory, permissions and an outbox. The Claude Platform best-practices article focuses on architecture. It explains that the persistent experience comes from repeated wake–work–sleep cycles: each period of work is a run, while a session holds the transcript of messages and tool calls. The underlying agent loop remains the same as in other agents.
The design addresses product teams making existing agents proactive or building background agents from scratch. Compared with agents triggered only by user messages, the event sources expand to webhooks, scheduled tasks and background jobs. The article warns that letting each source invoke the agent directly can cause competing runs or lost events after a crash. Its recommended design uses a durable inbox, permits only one active worker per inbox and batches pending events. Due timers also feed events into that inbox, while a separate durable memory store preserves state beyond the worker’s lifetime.
Unattended execution makes permission handling a decision at each tool call: proceed independently, block the action or wait for approval. Drawing on patterns used in Claude Code, the article describes an external proxy that injects credentials and a mechanism for deferring tool calls. When user confirmation is needed, the run ends and the session is saved mid-call; the user’s response returns through the inbox as an event that resumes execution. The worker can process other events while approval is pending. The article explicitly retains human review for the most important decisions, placing a boundary on autonomous operation.
Message delivery has a similarly bounded implementation choice. Anthropic says most always-on agents can use a tool that sends messages directly; a formal outbox and dispatcher are needed when delivery must be paced or recovered from errors. For the teams concerned, the engineering substance is the separation of event intake, persistent state, approval waits and delivery, extending beyond how the model performs a task. The article also notes that both compaction and starting a new session lose detail, so information that must survive belongs in memory during execution; notification frequency is something to tune through evaluations.
Evidence and limits
This edition relies on one official Anthropic engineering article. We did not independently verify its implementation or test deployment, recovery or permission security. The source does not establish availability as a complete managed product or demonstrate reliability and cost under a specific workload. No supporting material was supplied for market or regulatory coverage.
AI-assisted drafting with a separate model review. This is source-based analysis, not independent performance testing.