Agent autonomy needs a harness: layered state architecture survives a 10-day continuous run
burkov · x · 2026-11-01
Andriy Burkov highlights an article arguing that long-running research, software, and ops work outlives any model's context window, a single process, and the intervals when a human can supervise. Simply extending context doesn't fix it: state and constraints get lost on resets, goals drift, and unchecked errors compound.
The article proposes an architecture enabling continuous operation, validated in a ten-day test:
- Layered work organization: levels acting on different time scales, each keeping short, durable summaries
- A scheduled "tick" as the unit of autonomous action
- Explicit files recording instructions, decisions, and progress
- Task routing across models of different capabilities
Takeaway: the key to continuously operating agents isn't the model itself but the surrounding harness that schedules, stores, checks, and directs the work.
More from AGI Musings
- Nobel economist William Nordhaus dies; his work showed innovators capture just 2.2% of value — dscape · 2026-11-01
- David Manheim challenges LLM consciousness skeptic: why assume biology is mysterious but AI isn't? — davidmanheim · 2026-11-01
- Ex-YC interviewer who read ~10,000 applications warns AI writing erases founders' voice — amplifiedamp · 2026-11-01
- Anthropic reportedly believes Claude can go superhuman and is 'playing to win' — teortaxesTex · 2026-11-01
- Musk backs Page's prediction: Von Neumann probes could explore the galaxy by 2100 — heyshrutimishra · 2026-11-01
- Chatbots Are Leading Hikers Astray, With Dangerous Consequences — Wired AI · 2026-11-01