Running an always-on agent locally: 22 background runs used just 0.41% of quota across 1.51B tokens
Adorable-Cost-3249 · reddit · 2026-11-16
A Redditor implemented Anthropic's "Building always-on agents" guide locally without a cloud stack, using five primitives: a file-locked event inbox with 0-token change detection, persisted one-shot follow-up timers, duplicate-suppression memory (92% prefix-cache hits), bwrap/Seatbelt sandboxing with round-trip approvals, and a quiet-by-default outbox with a 10-case eval.
Across 186 real sessions (31,296 steps, 1.51B gross tokens): deep coding/research threads took 75.6% of quota, delegated subagents 19.8%, pinned chat 4.2% — while all 22 unattended background runs combined used just 0.41% (6.7M tokens, 45k fresh input/run, unchanged polls skipped at zero tokens). Ships as a single zero-dependency Go binary (Apache 2.0) with a 60-second demo sandbox.
More from coding & agent
- How are people handling agent versioning, memory, and cross-device sync? — yangyi · 2026-11-16
- Vibe-coded movie seat picker hits 200K users in 24 hours after 3-hour build — itsOmSarraf_ · 2026-11-16
- Proof of Agent hits 1M onchain transactions and 10,000 agents in just 27 days — adamamcbride · 2026-11-16
- Indie dev: Gemini4 Carbon matches 90-95% of Opus 5.5 for coding but is painfully slow — yihui_indie · 2026-11-16
- Who's building agents that recover from mistakes? Partial failures across systems are the blind spot — djgoel123 · 2026-11-16
- ICL alone may enable substantial continual learning, plot shows agents improve by Month 3 — xennygrimmato_ · 2026-11-16