Anthropic breaks down always-on agents: sleep-wake cycle and five core primitives
aigclink · x · 2026-11-17
Anthropic's new guide explains how to build always-on agents: they're mostly asleep, waking per event to run a receive-work-persist cycle — the "always on" feel is an architectural illusion.
Key points:
- Core loop is unchanged; the real engineering lives around it: when the agent wakes, what persists between runs, how work reaches users
- Five core primitives, starting with a persistent Event Inbox unifying webhooks, cron jobs, and user messages
- Resurgence driven by better model judgment and cheaper long-horizon intelligence: Opus 5.5 costs 78% less per solved task than Opus 5 on a long-horizon coding benchmark (vs 40% on typical workloads)
A practical reference for building digital employees and long-running agents.
Related event: Anthropic Unveils Architecture for Always-On AI Agents(3 posts)→
More from coding & agent
- Zed founder on agent skills: absurdly slow, quality all comes from babysitting — zeeg · 2026-11-17
- Dev open-sources inference-optimisation agent skill he says made him $100K — Kyrannio · 2026-11-17
- Open-source Hermes Agent profile: cheap main model plus low-cost side-task pinning — alexcovo_eth · 2026-11-17
- CUDA guide details fine-grained sync: sentinel-based tile readiness over membar, faster on Rubin — simran_s_arora · 2026-11-17
- Why Lean? Most momentum among dependent type languages, and agents write it decently — hargup13 · 2026-11-17
- Quarterdeck: open-source Claude Code plugin collapses Anthropic's always-on agent stack into one prompt — InternationalData569 · 2026-11-17