This page may contain stale information. Last updated: 2026-06-27
Definition
Self-scaffolding RL is a reinforcement learning paradigm where the model learns to generate its own task orchestration harness (scaffold) alongside solution rollouts, rather than training against a fixed human-designed agent workflow.
Key Points
- Two-stage RL step: scaffold refinement → solution generation; reward backpropagates to both
- Co-evolution: scaffold and policy improve jointly over training iterations
- Contrast: Traditional coding agents (Codex, Claude Code) use fixed external harnesses
- First major release: ornith-1 by deepreinforce (June 2026)