What Makes a Good Terminal-Agent Benchmark Task: A Guideline for Adversarial, Difficult, and Legible Evaluation Design

Forum topic · 小凯 · 2026-05-04

Summary

Terminal-agent benchmarks have become a primary signal for measuring the coding and system-administration capabilities of large language models, yet rapid task publishing often skips rigorous adversarial review. In this arXiv paper (2604.28093), Ivan Bercovich distills over a year of contributing to and reviewing Terminal Bench tasks into a practical guideline for writing good benchmark tasks. The core argument: most people write benchmark tasks the way they write prompts, but a prompt is designed to help an agent succeed while a benchmark is designed to test whether it can. The author proposes that good tasks are adversarial, difficult, and legible, and catalogues common failure modes including AI-generated instructions, over-prescriptive specifications, bureaucratic difficulty, oracle solutions that assume hidden knowledge, tests that validate incorrect content, and reward-hackable environments. The paper argues that genuine difficulty should be conceptual rather than environmental, and cites recent empirical evidence that over 15% of tasks in popular terminal-agent benchmarks are reward-hackable.

Paper Overview

Field: AI evaluation Author: Ivan Bercovich Published: 2026-04-30 arXiv: 2604.28093

Abstract

Terminal-agent benchmarks have become a primary signal for measuring coding and system-administration capabilities of large language models. This paper is a guideline for writing good benchmark tasks, drawn from over a year of contributing to and reviewing tasks for Terminal Bench.

Most people write benchmark tasks the way they write prompts - they shouldn't. A prompt is designed to help the agent succeed; a benchmark is designed to find out if it can.

Key Arguments

Good tasks are adversarial, difficult, and legible. These three properties form the guiding principle for benchmark task design.
Common failure modes include AI-generated instructions, over-prescriptive specifications, bureaucratic (paperwork) difficulty, oracle solutions that assume hidden knowledge, tests that validate incorrect content, and reward-hackable environments.
Root cause: These failure modes are predictable consequences of treating task-writing as prompt-writing.
Conceptual vs. environmental difficulty: Genuine difficulty should be conceptual rather than environmental.
Empirical evidence: Recent findings show over 15% of tasks in popular terminal-agent benchmarks are reward-hackable.

Conclusion

As the evaluation environment marketplace grows, the pressure to publish tasks quickly increases - often without thorough adversarial review of verification logic. This guideline aims to raise the quality bar for benchmark task authoring in the terminal-agent space.