AIAny
Icon for item

SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

Treats the interface between an LM agent and a computer as a design variable. A custom agent-computer interface (ACI) with concise file-edit, repo-navigation, and test commands plus compact feedback reaches 12.5% pass@1 on SWE-bench, 87.7% on HumanEvalFix.

Introduction

Plenty of work hands a language model the same shell and editor a human uses and assumes that is enough. This paper argues the opposite: agents are a distinct class of user, and the interface itself — what commands exist, how output is shaped — is the lever that moves performance. Just as IDEs were built for human ergonomics, agents need an agent-computer interface (ACI) designed for theirs.

Key Findings
  • The agent-computer interface (ACI) is treated as a first-class design surface: compact, purpose-built commands for navigating repos, viewing and editing files, and running tests, paired with feedback trimmed to what the model can actually use.
  • Interface design, not just the underlying model, drives results — SWE-agent reaches 12.5% pass@1 on SWE-bench and 87.7% on HumanEvalFix, far above non-interactive LM baselines.
  • Ablations show the shape of feedback and the granularity of edit commands materially change agent behavior, including how often it loops or recovers from mistakes.
Who It's For

Great fit if you build or study LM agents that act on real codebases and want evidence that interface engineering pays off as much as model choice. The SWE-agent framework and SWE-bench leaderboard make the work reproducible. Look elsewhere if you need a turnkey coding assistant rather than a research system, or if your interest is model training rather than the agent-environment boundary; the contribution is the interface and its analysis, not a new model.

Information

  • Websitearxiv.org
  • OrganizationsPrinceton Language and Intelligence, Princeton University
  • AuthorsJohn Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, Ofir Press
  • Published date2024/05/06

More Items

Analyzes adversarial weaknesses of World-Action Models (WAMs) via BadWAM, a framework that crafts visual perturbations to decouple a model’s imagined future from its executed actions. Introduces two attack modes—action-only (disruptive) and imagination-preserving (stealthy)—and shows large drops in closed-loop task success (e.g., 96.5%→43.1%).

Turns fragile, implicit search progress into explicit, persistent, shared state for multi-agent information seeking — externalizes progress as Frontier Task, Evidence Graph, Coverage Map and Failure Memory, and uses pipeline-parallel scheduling plus a middleware harness to avoid repeated failed searches and improve utilization and throughput.

Converts completed on-policy trajectories into natural-language 'hindsight skills' and converts the skill-induced action probability shifts into a dense token-level on-policy distillation signal, jointly optimized with outcome-based RL to improve sample efficiency and long-horizon agent behavior.