We need an L0.5 for continual learning: agent state stacks up in 4 layers
willcb · x · 2026-10-22
Seth Karten argues the current agent stack needs "an L0.5 for continual learning." The accompanying diagram lays out a four-layer state abstraction: L0 model weights (fine-tuning), L1 active context (compaction), L2 REPL & subagents with recursive session state (agentic garbage collection), and L3 disk-backed state holding history, artifacts, memories, skills, prompts, and subagent specs (refinement) — separated by a model-context boundary.
The point: between weights (L0) and context (L1) there's a missing layer that could carry genuine continual learning — an open zone for agent memory and self-improvement research.
More from coding & agent
- Shipping v5.0.0 with one prompt: Rork Max agent walks the whole app release — rudrank · 2026-10-22
- Claude Duet waitlist opens for US Pro/Max users with 6,000+ connectors — claudeai · 2026-10-22
- Claude Duet is built on Opus 5.5, inheriting memory, skills and apps — claudeai · 2026-10-22
- Claude Duet safety: no password access, new classifier guards sensitive data — claudeai · 2026-10-22
- Anthropic opens research preview of Claude Duet, a proactive AI agent app — claudeai · 2026-10-22
- A 12-layer engineering stack for production AI agents, from business to continual learning — MaryamMiradi · 2026-10-22