This arXiv preprint introduces OpenHands, a platform for building AI agents that act the way a human developer does. Its framing sets two things side by side: software as one of the most powerful tools humans have, which allows a skilled programmer to interact with the world in complex and profound ways; and the rapid development, thanks to improvements in large language models, of AI agents that interact with and affect change in their surrounding environments. The platform the authors propose joins them — agents that interact with the world in similar ways to those of a human developer, by writing code, interacting with a command line, and browsing the web.
The body of the paper is a description of what the platform provides. The authors describe how it allows for the implementation of new agents, safe interaction with sandboxed environments for code execution, coordination between multiple agents, and the incorporation of evaluation benchmarks.
On the strength of the benchmarks it has incorporated, the authors perform an evaluation of agents over 15 challenging tasks, which they state include software engineering (giving SWE-Bench as an example) and web browsing (giving WebArena as an example), among others.
The paper also states the project’s standing as software rather than only as a research artifact: OpenHands is released under the permissive MIT license and is a community project spanning academia and industry, with more than 2.1K contributions from over 188 contributors. On arXiv the preprint is filed under Software Engineering (cs.SE), with cross-listings to Artificial Intelligence (cs.AI) and Computation and Language (cs.CL). It was first submitted on 23 July 2024 and last revised on 18 April 2025 as version 3, carries the arXiv-issued DOI 10.48550/arXiv.2407.16741, and is noted as accepted by ICLR 2025. The code is stated to be at https://github.com/All-Hands-AI/OpenHands.
Key Points
- The paper introduces OpenHands, formerly known as OpenDevin, as a platform for developing powerful and flexible AI agents.
- Its design goal is agents that interact with the world in similar ways to a human developer — writing code, interacting with a command line, and browsing the web.
- The platform’s stated capabilities are implementing new agents, safe interaction with sandboxed environments for code execution, coordination between multiple agents, and incorporation of evaluation benchmarks.
- The authors evaluate agents over 15 challenging tasks drawn from the benchmarks the platform has incorporated, spanning software engineering and web browsing among others.
- OpenHands is released under the permissive MIT license, and the paper characterizes it as a community project spanning academia and industry with more than 2.1K contributions from over 188 contributors — counts that describe the project at the time of writing.
- The paper was accepted by ICLR 2025.
Notes
The paper’s contribution is a platform rather than a single agent, and its evaluation follows from that: the 15 tasks are described as being drawn from the benchmarks currently incorporated into the platform, so the reported coverage is a property of what has been integrated as much as of what the agents can do. The authors give the contribution and contributor counts as evidence of the project’s character as a community effort; both are point-in-time figures. The preprint went through three versions between July 2024 and April 2025.