THU-Agent Introduces Continual Harness Evolution Framework to Address Catastrophic Forgetting in Multi-Task Agents

Summary:The THU-Agent team at Tsinghua University has introduced a Continual Harness Evolution Framework to address the catastrophic forgetting problem in multi-task agents. Through a comprehensive empirical study involving three software engineering task datasets and six evolution directions, the research demonstrates that subsequent task learning can significantly degrade an agent's performance on previously seen tasks. For instance, Meta-Harness's performance on SWE-Bench dropped from 94.0% to 32.0%

---

Background and Motivation

In recent years, self-evolving agents have emerged as an important paradigm for enhancing agent performance. Prior studies have shown that evolving the agent harness can effectively improve agent capabilities without modifying the foundation model's parameters. However, existing research primarily focuses on agent evolution within a single task, while in practice, agents often need to handle multiple task types. Whether an agent harness can continuously evolve across multiple tasks without suffering from catastrophic forgetting remains an open question.

Methodology and Experiments

The study was conducted through the following steps:

1. Dataset and Task Design: The research team selected three software engineering task datasets and organized them into six different evolution directions.
2. Application of Evolution Methods: Two representative harness evolution methods (Meta-Harness and DeepSeek-V4-Flash-0731) were applied to perform continual evolution along these directions.
3. Performance Evaluation: The performance of the agents was evaluated on each task dataset.

The experimental results showed that subsequent task learning significantly degraded the agent's performance on previously seen tasks. For example, Meta-Harness's performance on SWE-Bench dropped from 94.0% to 32.0%, a 66% relative decrease, after evolving on FeatureBench. This phenomenon highlights the significant challenges faced by agents in a multi-task environment.

Findings and Implications

Based on the experimental results, the research team identified the following key findings:

• Catastrophic Forgetting is Prominent: Agents tend to forget previously learned knowledge when learning new tasks, leading to a significant drop in performance.
• Dependence on Evolution Direction: Different evolution directions have varying impacts on agent performance, with some directions causing more severe forgetting.
• Potential Optimization Directions: The study suggests that introducing memory mechanisms, task prioritization, and incremental learning could alleviate the catastrophic forgetting problem.

Future Directions

This research provides new insights and methods for the continual self-evolution of agents. Future research directions include:

• Developing more effective memory mechanisms to retain cross-task knowledge.
• Designing smarter task scheduling strategies to balance new task learning and old task knowledge retention.
• Exploring meta-learning-based evolution methods to achieve more efficient cross-task knowledge transfer.

Code and Data

The research team has made the code available at: https://github.com/THU-Agent/continual-harness-evolution.

Conclusion

Continual self-evolution is a key direction for agent development, but catastrophic forgetting in a multi-task environment remains a challenge. This study provides new experimental data and optimization directions, contributing to the advancement of agent technology.

---

Source: Hugging Face Daily Papers (2026-11-02)

— END —

Tags: #Agentic Follow #Continual Learning Follow #Catastrophic Forgetting Follow #Multi-Task Learning Follow #THU-Agent Follow