Voyager — automatic curriculum, skill library, iterative refinement
automatic curriculum
proposes next task from current capability gap
curiosity-driven, bottom-up exploration
"learn to craft iron pickaxe" comes before diamond
production analog:
gap review — what skills missing in library
weekly curriculum pass by human + agent pair
skill library
skill = named code + description + deps
versioned, retrievable, composable
dedup on write; tag filter on retrieve
2026 production analog:
Claude Agent SDK skills
skillkit registry; MCP resources
iterative prompting
execution error + env feedback + self-verify
rewrite the skill, bump version
v1 fails to craft; v2 adds stick dep
underpinnings:
Self-Refine (Lesson 05)
CRITIC (Lesson 05) with env verifier
worked loop — craft_iron_pickaxe
1. retrieve: library.search("need a pickaxe") returns craft_iron_pickaxe v1
2. execute v1: topo_order place_table mine_ore craft_iron_pickaxe
3. fail: RuntimeError "need 2 stick, have 0" environment feedback surface
4. refine: register v2 with gather_sticks dep; re-execute; success; inventory += iron_pickaxe
action space = code. skills compose into DAGs. failures bump versions. library grows monotonically.