Anthropic Breaks Down the Five Primitives Behind Always-On AI Agents
dotey · x · 2026-11-14
Anthropic published "Building always-on agents" (by William Hu), a guide to the primitives behind agents that run indefinitely across sessions. They're not literally always running—they sleep most of the time and wake on events, which keeps costs down.
- Inbox: events trigger a run, so the agent only wakes when something arrives
- Timer: lets it act proactively without external triggers
- Memory folder: carries state across runs; needs pruning to stay cheap
- Permissions: constrain what the agent can do with your accounts, with important actions awaiting user approval
- Outbox: decides when and how it speaks to the user
The post also covers practical concerns: handling concurrent events, why memory should be economical, and why new agents should start quiet.
The resurgence is driven by model improvements and cheaper intelligence: on one long-horizon coding benchmark, Opus 5.5 cost 78% less per solved task than Opus 5, versus 40% less on typical workloads—making long-horizon autonomous tasks economically viable.
More from coding & agent
- Dev open-sources React Dynamic Island component islandcn after thousands in AI inference costs — HankYeomans · 2026-11-14
- LangChain launches LangSmith platform: one-click agent deploy with 24/7 auto-monitoring — hwchase17 · 2026-11-14
- Tire Swing launches fully autonomous AI to automate affordable housing compliance — ycombinator · 2026-11-14
- 130K-follower KPI... 'Vibe coding as recreation': 15 minutes of prompting to train a model on his messages — gabriel1 · 2026-11-14
- Google's Antigravity multiagent proves 2013 info-theory conjecture in Lean 4 — ymatias · 2026-11-14
- Read-only MCP setup lets Claude Code audit a Proxmox lab — Worldly_Piglet_7040 · 2026-11-14