SCOPE: Prompt Evolution for Enhancing Agent Effectiveness

Abstract: LLM agents are increasingly deployed in environments that generate massive, dynamic contexts. However, a critical bottleneck remains: while agents have access to this context, their static prompts lack the mechanisms to manage it effectively, leading to recurring Corrective and Enhancement failures. To address this capability gap, we introduce \textbf{SCOPE} (Self-evolving Context Optimization via Prompt Evolution). SCOPE frames context management as an \textit{online optimization} problem, synthesizing guidelines from execution traces to automatically evolve the agent's prompt. We propose a Dual-Stream mechanism that balances tactical specificity (resolving immediate errors) with strategic generality (evolving long-term principles). Furthermore, we introduce Perspective-Driven Exploration to maximize strategy coverage, increasing the likelihood that the agent has the correct strategy for any given task. Experiments on the HLE benchmark show that SCOPE improves task success rates from 14.23\% to 38.64\% without human intervention. We make our code publicly available at https://github.com/JarvisPei/SCOPE.

• The paper introduces SCOPE, a self-evolving prompt framework that adapts agent behavior through online guideline synthesis.
• It employs dual-stream routing to classify and optimize tactical and strategic memory for effective context management.
• Experimental results show success rates rising from 14.23% to 38.64%, validating the framework's efficacy on GAIA and HLE benchmarks.

SCOPE: Prompt Evolution for Enhancing Agent Effectiveness

Introduction

The paper "SCOPE: Prompt Evolution for Enhancing Agent Effectiveness" (2512.15374) addresses a critical challenge in the domain of agentic systems powered by LLMs: the management of dynamic and complex contexts. As agents are increasingly tasked with solving multifaceted problems, they must navigate and process extensive information that often exceeds the capabilities of static prompts. This limitation results in corrective and enhancement failures, where agents either fail to adapt to errors or miss optimization opportunities. The paper introduces SCOPE (Self-evolving Context Optimization via Prompt Evolution), a framework designed to evolve agent prompts via online optimization, balancing tactical error correction with strategic principle enhancement.

Figure 1: Overview of the SCOPE Framework. The Generator ($\pi_\phi$) synthesizes candidate guidelines from execution traces. A Selector ($\pi_\sigma$) chooses the best guideline, which is then routed by the Classifier ($\pi_\gamma$) to either Tactical (task-specific) or Strategic (persistent) memory. The Optimizer ($\pi_\omega$) consolidates the strategic memory before updating the prompt.

Observations and Motivation

The investigation into agent failures under the GAIA and HLE benchmarks reveals two primary failure modes: corrective failures and enhancement failures. Corrective failures occur when agents treat error logs as generic alarms without leveraging the solutions embedded within these errors, leading to recursive error loops. Enhancement failures manifest when agents persist with suboptimal strategies, ignoring patterns that indicate potential optimization paths.

The complexity inherent in agentic systems—comprising specialized agents like coders, analyzers, and browsers—magnifies these failures. Yet, this complexity also presents a unique opportunity: repeated agent invocation within a system allows for continuous learning and prompt evolution aligned with context management needs. SCOPE leverages this structural opportunity to transform these failures into insights through guideline synthesis from execution traces, enabling both corrective and enhancement adjustments.

Methodology

SCOPE's methodology fundamentally transforms context management from a manual task into an automatic optimization process. The framework operates through four components:

1. Guideline Synthesis: Utilizing execution traces to generate corrective and enhancement guidelines via a generator ($\pi_\phi$) and selecting the most promising guideline through a selector ($\pi_\sigma$).
2. Dual-Stream Routing: Classifying guidelines into tactical (task-specific) and strategic (long-term principle) streams using a classifier ($\pi_\gamma$), allowing for tailored adaptability.
3. Memory Optimization: Consolidating strategic memory through an optimizer ($\pi_\omega$), ensuring guideline quality and coherence by merging similar guidelines, resolving conflicts, and pruning redundant guidelines.
4. Perspective-Driven Exploration: Implementing multiple parallel optimization streams with distinct perspectives (e.g., efficiency vs. thoroughness) to maximize strategy coverage across diverse tasks.

   Figure 2: Performance breakdown by subcategory on HLE (a) and GAIA (b). SCOPE performs consistently across domains, with notable gains in complex protocol adherence.

Experimental Results

Experiments demonstrate that SCOPE significantly enhances task success rates on the HLE and GAIA benchmarks, outperforming static agent configurations and existing optimization methods such as DC and ACE. On expert-level tasks, SCOPE's prompt evolution mechanism raises success rates from 14.23\% to 38.64\%, showcasing its efficacy in dynamic environments where context management is critical.

This improvement is driven by the framework's ability to integrate learned guidelines directly into agent prompts, adapting mid-task to recover from errors and optimize strategy. The ablation study highlights that each component contributes effectively, with Perspective-Driven Exploration yielding the largest performance gain by embracing diverse optimization strategies.

Implications for AI Development

SCOPE's approach to prompt evolution offers several theoretical and practical implications:

• Theoretical Implications: Enables a shift from static prompt engineering to dynamic evolution, where agents continually refine their strategies based on execution feedback.
• Practical Implications: Opens avenues for deploying adaptive agents in real-world tasks, reducing error rates while enhancing strategic execution across heterogeneous domains.

Future developments in AI could explore more granular adaptation mechanisms, refined guideline classification strategies, and enhanced memory optimization techniques to further expand the utility of self-evolving agent systems.

Conclusion

SCOPE defines a new paradigm in AI agent design, integrating prompt evolution to address context management challenges in dynamic environments. By empowering agents with the capability to synthesize and integrate execution-based guidelines, SCOPE advances the field beyond traditional static prompts, paving the way for more adaptive, effective agentic systems.