Anthropic Breaks Down the 5 Core Components of Building Always-On AI Agents
aigclink · x · 2026-11-17
Anthropic explains how to build resident (always-on) AI agents: the agent mostly "sleeps," waking only on event triggers to run a task, write results to persistent storage, and sleep again — the sense of always being present is an engineered illusion. The hard engineering lies outside the main loop, in five core components:
- Event Inbox: a persistent entry point unifying webhooks, cron jobs, and user messages, processed by a worker in order to avoid loss, duplication, and concurrency conflicts.
- Timers: giving agents a sense of time for follow-ups and recurring tasks ("check last night's build at 9am," "weekly report every Friday").
- Memory: stored independently of sessions in persistent storage; each run loads instructions and an index first, then fetches details on demand. Without it, every wake-up is an amnesiac restart.
- Permissions: credential proxies, auto modes, deferred tool calls — defining what the agent may do autonomously vs. what needs human approval. This is the toy-to-production watershed.
- Outbox: reliable delivery of results, reminders, and approval requests, with notification pacing and retry handling.
Key design: memory and sessions are decoupled — sessions are ephemeral, memory is durable, which distinguishes stateful agents from stateless chatbots. Full loop: trigger → inbox → worker wakes agent → load memory → execute → save state → notify if needed → sleep.
Related event: Anthropic Details Architecture of Always-On Agents(3 posts)→
More from coding & agent
- Vibecoder rebuilds entire world in Minecraft at 1:1 scale with 250+ AI agents — Polymarket · 2026-11-17
- Sentry founder on pstack verification skills: it's basically QA skills, prescriptive but useful — zeeg · 2026-11-17
- I had Claude, Codex, Grok and Deepseek build games from 15 years of old Humble Bundle assets — slickriptide · 2026-11-17
- Kosem launches: coding agents with native environments, identity and access — jekapats · 2026-11-17
- Dev shows how he talks to AI: a hand-drawn top-down scene sketch as the prompt — ZeroStateReflex · 2026-11-17
- Charlie Holtz shows off a GUI for his coding harness-harness orchestrator — charlieholtz · 2026-11-17