# VeRO: A Harness for Agents to Optimize Agents

**Authors:** Varun Ursekar, Apaar Shanker, Veronica Chatrath, Yuan Xue, Samuel Marc Denton  
**arXiv:** [2602.22480](https://arxiv.org/abs/2602.22480)  
**Date:** 2026-02-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.22480v4/

## Key points

- Study explores applicability of VeRO in optimizing agents for complex and long-horizon coding tasks
- Case study conducted using TerminalBench-2 benchmark with Terminus-KIRA agent and Claude Haiku 4.5 LLM
- Two modes tested: Tools interface and Filesystem interface for exposing execution traces and dataset content to optimizer
- Three optimization runs conducted using Claude Code (Sonnet 4.5) with varying sample budgets under different interfaces
- Results show improvements in pass rates over baseline agent, identifying fixes that enhance performance
- Nuanced dynamics observed with different combinations of fixes found in each run, highlighting complexity of optimizing agent harnesses
- Interpretability explored through Git commit histories to investigate semantic trends in optimization process for various tasks
- GPT-4.1 used to tag changes made by coding agent during each optimization trajectory, providing insights into impact on agent performance
- Study emphasizes effectiveness of VeRO in optimizing agents for long-horizon coding tasks and importance of interpretability in understanding optimization processes

## Summary

This study explores the applicability of VeRO in optimizing agents for complex and long-horizon coding tasks. A case study is conducted using TerminalBench-2, a benchmark consisting of 89 terminal tasks evaluated in sandboxed containers. The base target agent used is Terminus-KIRA with Claude Haiku 4.5 as the underlying LLM. Two modes of exposing execution traces and dataset content to the optimizer are tested: Tools interface and Filesystem interface. Three optimization runs are conducted using Claude Code (Sonnet 4.5) as the optimizer with varying sample budgets under different interfaces. Results show improvements in pass rates over the baseline Terminus-KIRA agent, with both Tools and Filesystem interfaces identifying fixes that enhance performance. However, nuanced dynamics are observed where different runs find different combinations of fixes, indicating the complexity of optimizing agent harnesses. Interpretability is also explored by leveraging Git commit histories to investigate semantic trends in the optimization process for various tasks. GPT-4.1 is used to tag changes made by the coding agent during each optimization trajectory, providing insights into how optimizations impact agent performance. Overall, this study highlights the effectiveness of VeRO in optimizing agents for long-horizon coding tasks and underscores the importance of interpretability in understanding optimization processes. The findings contribute to advancing research on agent optimization as a crucial capability for coding agents.