# Rethinking Self-Evolving Agents: Do We Still Need Prescribed Optimization Pipelines?

> Frontier optimizers can compose task-specific improvement strategies online without prescribed pipelines, matching or beating them on 12 of 14 settings while using a third of the compute.

- **Source:** [arXiv](https://arxiv.org/abs/2608.09629)
- **Published:** 2026-08-22
- **Permalink:** https://picx.dev/p/63a00p
- **Whiteboard:** https://picx.dev/p/63a00p/image

## Summary

# Summary of "Rethinking Self-Evolving Agents: Do We Still Need Prescribed Optimization Pipelines?"

## Summary (Overview)

- **Core Question**: The paper investigates whether task-specific prescribed optimization pipelines are still necessary for self-evolving agents when a frontier model (GPT-5.5) acts as the optimizer, or whether the optimizer can compose the improvement process online.

- **Novel Framework**: The authors introduce **Open-Ended Optimization (OEO)**, which fixes the external "optimization contract" (objective, permissions, budget, data boundaries, evaluation) but allows the optimizer to compose the task-specific "optimization meta-policy" (evidence gathering, revision strategy, stopping criteria) online.

- **Key Results**: Across 14 head-to-head comparisons over 8 benchmark–target-model settings, GPT-5.5-driven OEO records **12 wins, 1 tie, and 1 narrow loss of 0.21 percentage points**, while using a median **34.3%** of SKILLOPT's configured target-interaction token budget.

- **Control Experiments**: A one-shot, zero-interaction rewrite control shows gains are not explained by prior-driven rewriting alone; however, delegation has a capability boundary—SKILLOPT outperforms OEO with a medium optimizer, and a weak optimizer cannot operate through the unchanged OEO interface.

- **Process Analysis**: Trajectory analysis reveals that prescription changes *how* optimization proceeds more consistently than it changes final behavior, with distinct paths often converging on overlapping item-level outcomes.

## Introduction and Theoretical Foundation

### Background

Self-evolving agents aim to turn experience into persistent improvement by revising prompts, memories, skills, or tool-use policies. Most existing systems prescribe not only *what* should improve but *how* improvement should happen—including which evidence to collect, how to diagnose failures, what revisions to propose, which candidates to retain, and when to stop.

**The core distinction introduced** separates:
- **Optimization contract (C)**: The externally governed framework that fixes the objective, initial skill $s_0$, target model, permitted selection data $D_{sel}$, permitted task-facing operations $\mathcal{O}$, resource budget $B$, frozen evaluator, and sealed final split $D_{final}$.
- **Optimization meta-policy**: The task-specific logic organizing evidence, revision, selection, and stopping—which the framework may prescribe or the optimizer may compose online.

### Formal Framework

Given state $z_t = (s_t, H_t, K_t, b_t)$ (current skill, observations, committed checkpoints, remaining resources), the optimization meta-policy selects:

$$
a_t = \mu^{(q)}(z_t; M_{\mathrm{opt}}, \mathcal{C}, \mathcal{I}_q), \qquad a_t \in \mathcal{A}_q(\mathcal{O}), \tag{1}
$$

where $\mathcal{A}_q(\mathcal{O})$ is the procedure-specific representation of permitted operations.

### Two Representative Prescribed Approaches

The paper compares OEO against two structurally different prescribed procedures:

| Control Dimension | SKILLOPT | GEPA | OEO |
|---|---|---|---|
| Evidence scheduling | Fixed rollout/reflection batches | Parent and minibatch selection | Optimizer-directed |
| Diagnosis | Success/failure decomposition | Trace-conditioned reflection | Optimizer-directed |
| Revision space | Bounded patch operators | Free-form textual mutation | Optimizer-selected edit/rewrite |
| Candidate retention | Strict improvement gate | Pareto-aware pool update | Optimizer selects among committed states |
| Schedule/stopping | Fixed stages and epochs | Budgeted evolutionary loop | Adaptive within budget |

- **SKILLOPT**: Pipeline-centric prescription with staged rollout, bounded skill edits, and validation-gated updates.
- **GEPA**: Search-centric prescription with free-form textual mutation inside a Pareto-aware reflective evolutionary loop.

## Methodology

### OEO Protocol

OEO is a contract-enforced agent loop where at each turn the optimizer observes the current skill, accumulated evidence, committed checkpoints, and remaining resources, then selects one permitted operation. The runner validates and executes the operation, records results and costs, and prevents access to the sealed final split. Skills are committed immediately for subsequent rollouts, and the optimizer may select any committed checkpoint as its final output.

### Experimental Setup

- **Benchmarks**: SearchQA (1,400 items), SpreadsheetBench (280 items), OfficeQA (172 items), LiveMathematicianBench (124 items)
- **Target models**: Qwen3.5-4B and GPT-5.5
- **Optimizer**: GPT-5.5 for all methods in the frontier comparison
- **Evaluation**: Sealed final splits with frozen evaluators

### Control Conditions

1. **Static-input-matched one-shot control**: One GPT-5.5 rewrite with zero target-model rollouts
2. **Capability ladder**: Weak (Qwen3.5-4B), medium (Qwen3.5-27B), and frontier (GPT-5.5) optimizers with frozen protocols

### Trajectory Diagnostics

The fully instrumented OEO–SKILLOPT pair was analyzed using:
- **Path measures**: Maximum normalized edit, section breadth, revision churn
- **Outcome measures**: Pass/fail agreement, correct-set Jaccard ($J_C$), gain-set Jaccard ($J_G$)

## Empirical Validation / Results

### Frontier Comparison (Table 2)

| Benchmark | Target | Initial | SKILLOPT | GEPA | OEO |
|---|---|---|---|---|---|
| SearchQA | Qwen3.5-4B | 0.6693 | 0.7221 | 0.7507 | **0.7514** |
| SearchQA | GPT-5.5 | 0.7857 | 0.8657 | **0.8736** | 0.8714 |
| SpreadsheetBench | Qwen3.5-4B | 0.2107 | 0.1964 | 0.2214† | **0.2464** |
| SpreadsheetBench | GPT-5.5 | 0.3964 | **0.7607** | 0.4964† | **0.7607** |
| OfficeQA | Qwen3.5-4B | 0.2616 | 0.2849 | 0.2500 | **0.3081** |
| OfficeQA | GPT-5.5 | 0.5872 | 0.7035 | 0.6977 | **0.7267** |
| LiveMath | Qwen3.5-4B | 0.3065 | 0.6129 | 0.6210 | **0.6532** |
| LiveMath | GPT-5.5 | 0.3871 | 0.5000 | 0.4032 | **0.6855** |

**Key findings**: OEO improves all 8 initial skills, SKILLOPT improves 7, and GEPA improves 5 of its 6 confirmatory settings. The largest margin is +18.55 points on GPT-5.5-target LiveMath.

### One-Shot Control (Table 3)

| Benchmark | Target | Initial | One-shot rewrite | OEO |
|---|---|---|---|---|
| SearchQA | Qwen3.5-4B | 0.6693 | 0.6900 | 0.7514 |
| SearchQA | GPT-5.5 | 0.7857 | 0.8357 | 0.8714 |
| LiveMath | Qwen3.5-4B | 0.3065 | 0.2742 | 0.6532 |
| LiveMath | GPT-5.5 | 0.3871 | 0.3468 | 0.6855 |

Static prior helps on SearchQA but *regresses* on LiveMath, while OEO improves both—demonstrating the interactive loop supplies gains a single rewrite cannot.

### Capability Ladder (Table 4)

| Benchmark | Method | Weak Qwen3.5-4B | Medium Qwen3.5-27B | Frontier GPT-5.5 |
|---|---|---|---|---|
| LiveMath | OEO | blocked | 0.4919 | 0.6532 |
| LiveMath | SKILLOPT | 0.5565 | 0.5887 | 0.6129 |
| SearchQA | OEO | blocked | 0.6807 | 0.7514 |
| SearchQA | SKILLOPT | 0.6579 | 0.7157 | 0.7221 |

The crossover is clear: at medium capability, SKILLOPT leads by 9.68 points (LiveMath) and 3.50 points (SearchQA). The weak optimizer cannot produce executable actions through the OEO interface.

### Trajectory Analysis (Table 5)

| Benchmark | Target | OEO | SKILL | Agreement | $J_C$ | $J_G$ |
|---|---|---|---|---|---|---|
| SearchQA | Qwen3.5-4B | 0.7514 | 0.7221 | 0.8793 | 0.8486 | 0.5567 |
| SearchQA | GPT-5.5 | 0.8714 | 0.8657 | 0.9629 | 0.9581 | 0.7500 |
| SpreadsheetBench | Qwen3.5-4B | 0.2464 | 0.1964 | 0.8000 | 0.3778 | 0.2653 |
| SpreadsheetBench | GPT-5.5 | 0.7607 | 0.7607 | 0.9000 | 0.8767 | 0.8136 |
| OfficeQA | Qwen3.5-4B | 0.3081 | 0.2849 | 0.8023 | 0.5000 | 0.3000 |
| OfficeQA | GPT-5.5 | 0.7267 | 0.7035 | 0.9070 | 0.8779 | 0.7273 |
| LiveMath | Qwen3.5-4B | 0.6532 | 0.6129 | 0.7823 | 0.7065 | 0.7500 |
| LiveMath | GPT-5.5 | 0.6855 | 0.5000 | 0.7016 | 0.5978 | 0.3617 |

**Key insight**: OEO makes broader edits and exhibits higher revision churn in all 8 settings, yet pass/fail agreement exceeds 0.78 in 7 settings and correct-set Jaccard exceeds 0.70 in 5. The descriptive Spearman correlation between final-text distance and behavioral distance is only $\rho = 0.3095$—textual divergence does not imply behavioral divergence.

## Theoretical and Practical Implications

1. **Capability-adaptive division of labor**: Prescribed pipelines are repositioned as capability-dependent scaffolding rather than the sole source of optimization intelligence. At frontier capability, delegation is competitive; at lower capability, prescription becomes advantageous.

2. **Framework governance remains essential**: Delegation does not remove governance. Objectives, permissions, budgets, evaluation, and data boundaries must always remain framework-owned—only the task-specific route to improvement moves to the optimizer.

3. **Process and function require separate evaluation**: Skill text, committed trajectory, and evaluated behavior are distinct objects. Neither a text diff nor a final benchmark score alone explains what the system learned. Similar aggregate scores can conceal complementary correct-item sets, while different paths can reach overlapping behavior.

4. **Design implications**: The paper suggests a capability-adaptive design principle: begin with delegated composition when the optimizer can use executable feedback effectively, and introduce stronger procedural support when it cannot reliably produce or select useful updates.

## Conclusion

The paper demonstrates that **a sufficient optimizer can compose the route from measurable feedback to persistent improvement** without inheriting a prescribed pipeline. The central principle is **capability-adaptive responsibility**: keep the optimization contract external, treat task-specific prescription as a design choice rather than a default when progress is measurable, and delegate online composition only to an optimizer that can carry it effectively.

**Future directions** implied by this work include:
- Identifying principled criteria for when to delegate vs. prescribe the meta-policy
- Developing adaptive systems that can transition between delegated and prescribed modes based on measured optimizer capability
- Extending the OEO framework to other persistent artifacts (prompts, memory, tool-use policies) beyond natural-language skills
- Developing process-level diagnostics that complement item-level benchmark evaluation for self-evolving systems

---

_Markdown view of https://picx.dev/p/63a00p, served by PicX — AI-generated visual whiteboard summaries of research papers._