OEO Lets GPT‑5.5 Orchestrate Its Own Agent Improvement Process, Winning 12 of 14 Comparisons

Self-improving agents typically follow a framework-prescribed process: execute tasks, collect failure trajectories, reflect, propose skill or prompt modifications, and then use a validation set to determine whether to accept them. The newly proposed Open-Ended Optimization (OEO) retains external constraints while allowing the optimizer model to decide in real time how to combine these steps. Fixed elements include the task objective, permitted interactions, resource budget, data boundaries, and final evaluation; the model can decide when to analyze trajectories, test hypotheses, rewrite persistent artifacts, or stop searching.

The study compared OEO with two prescriptive approaches. SkillOpt trains Markdown skills using rollouts, reflection, constrained add/delete/replace edits, and a holdout validation gate. GEPA evolves prompts through natural-language reflection, candidate mutations, and Pareto selection. Across 14 head-to-head comparisons spanning eight benchmark and target-model configurations, GPT‑5.5-driven OEO recorded 12 wins, one tie, and one loss by just 0.21 percentage points. Its median target-agent interaction token usage was only 34.3% of the budget configured for SkillOpt. A one-shot, zero-interaction rewriting control failed to explain these gains, suggesting that the effect was not merely a strong model using its prior knowledge to rewrite a better skill.

These results do not mean that every agent harness can be removed. OEO still depends on permissions, budgets, data isolation, and evaluators that the model cannot change; what is open-ended is the optimization strategy, not the safety boundary. The experiments also revealed a clear capability threshold: when a medium-strength optimizer was substituted, SkillOpt came out ahead, while weaker models could not conduct an effective search through the same interface. In other words, as model capabilities improve, fixed workflows may shift from necessary algorithms to scaffolding, but validation gates and governance layers still cannot be entrusted to the same self-modifying loop.

Engineering teams should next watch for two things: whether OEO releases a reproducible implementation and complete trajectories, and whether its advantage persists across models from different providers, noisy evaluators, and longer-term skill evolution. The current evidence covers only 14 comparisons, and the core advantage depends heavily on GPT‑5.5. Until public code, detailed cost breakdowns, and more replicated trials are available, the 34.3% token figure should be treated as relative usage under a specific experimental configuration, not a general service-cost commitment.