Agentless challenges the prevailing assumption that autonomous agents are the best approach to AI-powered code repair. The system uses a two-phase pipeline: localization identifies the relevant files, classes, and functions through hierarchical search, then repair generates patches using the narrowed context. This structured approach avoids the unpredictability of agent loops where an AI repeatedly tries different tools and strategies with escalating costs.
The results are striking: on SWE-Bench Lite, Agentless achieves competitive scores with top agent-based systems while averaging just $0.34 per issue — a fraction of the cost of approaches like Devin or OpenHands that may spend dollars per attempt. OpenAI adopted the Agentless approach for evaluating o3's software engineering capabilities, lending significant credibility to the methodology. The simplicity also means fewer failure modes and more predictable behavior.
Agentless is MIT licensed with 3,000+ GitHub stars and active research development. It supports multiple LLM backends and works on real-world GitHub issues. For teams evaluating AI coding tools, Agentless represents an important design philosophy: sometimes a well-structured pipeline outperforms autonomous agents, especially when cost efficiency and predictability matter more than handling the most complex edge cases.
