Long-Horizon Harnesses: A Nine-Layer Architecture for Agents That Run for Months
KadriJibraan · x · 2026-11-19
- muhib waqar published a 22-min article on long-horizon harnesses: the software around a model (tools, memory, permissions, loop) that keeps an agent working toward one goal for months after the chat ends.
- Context: METR finds frontier agent task length doubling every 7 months, but that measures hours-long human tasks, while health and money run on months.
- Covers existing practices: Karpathy's autoresearch (overnight ML experiments), Claude Code's /loop and cron tools, Rubric Labs' always-on agents (Instinct, Muse, Duet) woken by heartbeats and world events.
- Includes a nine-layer harness architecture, a copyable design, case studies in agent businesses/medtech/fintech (Anthropic, Replit, Cursor, Cognition, Abridge), plus the biggest objection and open questions.
More from coding & agent
- Lunch-break demo: ChatGPT + CapCut plugin turns camera roll photos into a GTA 6-style intro video — alifcoder · 2026-11-19
- Tensor launches as an ADE for doing real work with AI, live on Mac, Windows and web — saheedniyi_02 · 2026-11-19
- Ultrafast mode review: great for math proofs, not recommended for coding — gnukeith · 2026-11-19
- Speed up CI for AI-written code: caching, faster runners, test sharding — rseroter · 2026-11-19
- Agentcard raises $8M to let agents buy from 5,000+ stores via Stripe Link — jeff_weinstein · 2026-11-19
- hstack open-sources one-command self-hosting for the Hermes agent — TheMoonMidas · 2026-11-19