CITATION — REFERENCE ENTRY
OpenHands: An Open Platform for AI Software Developers as Generalist Agents — The Thirteenth International Conference on Learning Representations (ICLR 2025)
- Key
- wang2025openhands
- Authors
- Wang, Xingyao; Li, Boxuan; Song, Yufan; Xu, Frank F.; Tang, Xiangru; Zhuge, Mingchen; Pan, Jiayi; Song, Yueqi; Li, Bowen; Singh, Jaskirat; Tran, Hoang H.; Li, Fuqiang; Ma, Ren; Zheng, Mingzhang; Qian, Bill; Shao, Yanjun; Muennighoff, Niklas; Zhang, Yizhe; Hui, Binyuan; Lin, Junyang; Brennan, Robert; Peng, Hao; Ji, Heng; Neubig, Graham
- Issued
- 2025
- Type
- paper-conference
- Container
- The Thirteenth International Conference on Learning Representations (ICLR 2025)
- Pages
- 64910-64947
Raw CSL JSON
{
"URL": "https://arxiv.org/abs/2407.16741",
"note": "arXiv:2407.16741v3 [cs.SE], 18 April 2025. Version cited: v3.",
"page": "64910-64947",
"type": "paper-conference",
"title": "OpenHands: An Open Platform for AI Software Developers as Generalist Agents",
"author": [
{
"given": "Xingyao",
"family": "Wang"
},
{
"given": "Boxuan",
"family": "Li"
},
{
"given": "Yufan",
"family": "Song"
},
{
"given": "Frank F.",
"family": "Xu"
},
{
"given": "Xiangru",
"family": "Tang"
},
{
"given": "Mingchen",
"family": "Zhuge"
},
{
"given": "Jiayi",
"family": "Pan"
},
{
"given": "Yueqi",
"family": "Song"
},
{
"given": "Bowen",
"family": "Li"
},
{
"given": "Jaskirat",
"family": "Singh"
},
{
"given": "Hoang H.",
"family": "Tran"
},
{
"given": "Fuqiang",
"family": "Li"
},
{
"given": "Ren",
"family": "Ma"
},
{
"given": "Mingzhang",
"family": "Zheng"
},
{
"given": "Bill",
"family": "Qian"
},
{
"given": "Yanjun",
"family": "Shao"
},
{
"given": "Niklas",
"family": "Muennighoff"
},
{
"given": "Yizhe",
"family": "Zhang"
},
{
"given": "Binyuan",
"family": "Hui"
},
{
"given": "Junyang",
"family": "Lin"
},
{
"given": "Robert",
"family": "Brennan"
},
{
"given": "Hao",
"family": "Peng"
},
{
"given": "Heng",
"family": "Ji"
},
{
"given": "Graham",
"family": "Neubig"
}
],
"issued": {
"date-parts": [
[
2025
]
]
},
"accessed": {
"date-parts": [
[
2026,
8,
29
]
]
},
"container-title": "The Thirteenth International Conference on Learning Representations (ICLR 2025)"
}
Claims
-
OpenHands connects an agent to its environment through a small set of general actions: executing Python code in an interactive interpreter, running bash commands in a shell, and interacting with a web browser using a domain-specific language borrowed from BrowserGym. The action space is described as based on programming languages, which the authors argue is flexible enough to cover tasks otherwise handled by purpose-built tools.
"Actions ... enable the agent to execute arbitrary Python code and bash commands inside the sandbox"
-
A footnote states that the project was initially inspired by Devin, the AI software engineer built by Cognition AI, and that OpenHands subsequently broadened beyond software engineering through community contributions.
"While initially inspired by AI software engineer Devin (Cognition.ai), OpenHands has quickly evolved"
-
The authors' ethics statement acknowledges that as agents are deployed more widely they pose significant security risks alongside productivity gains, and argues OpenHands mitigates these by enabling systematic evaluation before wide deployment, by facilitating human-agent interaction rather than letting agents operate autonomously without oversight, and by giving researchers access to agents for safety research.
"Facilitating human-agent interaction rather than allowing agents to operate autonomously without oversight"
-
OpenHands is released under the MIT license, which permits commercial use. At the time of writing the paper reported 32,000 GitHub stars and more than 2,100 contributions from over 188 contributors, and described the project as spanning academia and industry.
"Released under the permissive MIT license ... more than 2.1K contributions from over 188 contributors"
-
For each task session OpenHands starts an isolated Docker container. Inside it runs an action execution API server that maintains a bash shell, a Jupyter IPython server for interactive Python execution, and a Chromium browser driven by Playwright. A user-specified workspace directory is mounted into the container. OpenHands can install this API server into an arbitrary user-supplied Docker image.
"For each task session, OpenHands spins up a securely isolated docker container sandbox"
-
On SWE-Bench Lite, a 300-instance subset, CodeActAgent v1.8 resolved 26.0% of issues using claude-3-5-sonnet at an average cost of $1.10 per instance, 22.0% using gpt-4o at $1.72, and 7.0% using gpt-4o-mini at $0.01. Results were obtained without the benchmark's optional hint text. The paper notes that running the full 2,294-instance SWE-Bench set was estimated to cost about $6,900.
"CodeActAgent v1.8, using claude-3.5-sonnet, achieves a competitive resolve rate of 26%"
Available in