An experiment: can an agent's harness — not its plan, not its code — be an emergent, adapting artifact that improves across runs?
Assembled from the bare suite (bareagent, litectx, bareguard); builds no primitives.
A deliberately dumb outer loop holds the grounded close and the budget; the agent authors its
own workflow as a constrained config, revises it mid-run when the close fails, and inherits
what worked (verdict-gated) across runs. The agent never authors its arbiter.
Sibling of relayfact, which settled the grounded-loop question this experiment stands on. Like relayfact, adaptlearn either graduates or gets archived — the bar is the point.
The governing spec: docs/01-product/adaptlearn-prd.md (v1, locked). Status: pre-build — module ladder M0–M6 defined, nothing built yet.
Apache-2.0.