Self-Developing Agents: Fully Closed-Loop Recursive Self-Improvement Fails Without External Teacher

ByteDance Seed and TokenWave introduced Self-Developing Agents, a framework to test fully closed-loop recursive self-improvement (RSI) without an external verifier. Three benchmarks were designed: Aspire (goal formation from vague objectives), S³Gym (experience integration across 7 games with self-testing/judging/improvement), and HarnessDev (building and evolving an agent harness). Key findings: goal formation changed search behavior but not capability; no universal experience path existed across games; self-judging could not predict improvement gains (ρ ≈ −0.01); harnesses contained dead mechanisms; only 2 of 9 final versions were optimal on held-out evaluations. The work argues for testing RSI on unseen held-out tasks and using fixed executors to verify genuine improvement.