Anthropic details always-on agent architecture: inbox, timers, memory, permissions
AI寒武纪 · wechat · 2026-11-16
An Anthropic engineering article reveals the full architecture behind "always-on" autonomous agents that spend most of their time dormant—consuming no tokens—and wake only when triggered, chaining short runs into what feels like 24/7 availability.
Why this is exploding now: newer models judge intent well enough to know when to report back, and long-task costs collapsed—Opus 5.5 solves individual tasks 78% cheaper than Opus 5, with routine tasks down 40%.
The five building blocks:
- Event inbox: a DB table + HTTP endpoint + worker service as the single source of truth; batches consecutive messages, one active worker per inbox.
- Timers: follow-up alarms the agent sets itself, plus human-scheduled recurring tasks, polled by a scheduler that pushes events into the inbox.
- External memory: a persisted directory storing only what can't be looked up elsewhere, split into user instructions, notes, sent-message history, and task progress; a background "dreaming" process prunes stale records, with prompt caching and summarization controlling context costs.
- Permissions: three layers—a credential proxy so no secrets live inside the agent, a built-in safety model blocking risky actions, and deferred tool calls that checkpoint state and wait for human approval cards.
- Outbox: all outbound messages queue in a DB table for a dispatch service; best practice is staying quiet early and earning attention as accuracy improves.
More from coding & agent
- Gemini 4 Hydrogen launched: 2.5x faster than 3.8 Flash, but how much is batching? — krishnan · 2026-11-16
- Running 13 agents in parallel: handhold the critical bits, let the rest rip all day — willcb · 2026-11-16
- Agent tool calls stall silently with no worker polling, warns deployment thread — krishnan · 2026-11-16
- Agent sessions stall silently when no worker polls—ops must watch for it — krishnan · 2026-11-16
- Self-hosting agents buys control but you own the server, database, sandboxes and every release — krishnan · 2026-11-16
- A test worth requesting for agent platforms: kill a worker mid-job and see what recovers — krishnan · 2026-11-16