OpenHands: An Open Platform for AI Software Developers as Generalist Agents research paper summary (October 2026)

The OpenHands paper, accepted at ICLR 2025, describes an open platform for building AI agents that act like software developers. Its repository has since grown into a self-hosted control center for coding agents that also runs third-party agents.

The paper introduces OpenHands, formerly OpenDevin, a platform where agents interact with the world by writing code, using a command line and browsing the web. It covers safe interaction with sandboxed environments, coordination between multiple agents and the integration of evaluation benchmarks.

Today the repository presents Agent Canvas, a self-hosted developer control center that runs the OpenHands agent and ACP-compatible agents such as Claude Code or Codex locally, on servers or on OpenHands Cloud. It is under the MIT license, had 89,367 stars and was updated on September 28, 2026.

• Describes a reference agent architectureThe paper explains how an agent platform separates the model, the sandbox and the tools. Architects designing internal agent systems get a documented pattern to compare vendor offerings against.
• Runs coding agents on own infrastructureThe current Agent Canvas runs agents in Docker containers, on VMs or inside company infrastructure. Teams that cannot send code to a vendor cloud can operate agents themselves with a model of their choice.
• Automates routine engineering tasksAutomations trigger agents on schedules or webhook events, for example to break GitHub issues into tasks or post reports to Slack. Engineering teams hand off repetitive chores while keeping review in their own hands.

• Peer-reviewed at ICLR 2025 with open code under the MIT license.
• Self-hosting is the default, with optional cloud and enterprise offerings from the project.
• Also runs third-party agents through the Agent Client Protocol.

The paper is freely available on arXiv, and the accompanying code runs on your own hardware in any country.

The repository today differs substantially from the system described in the paper, so read the current documentation. Self-hosting keeps code in your environment, but the language model provider you connect still receives prompts and code. Follow the self-hosting security guidance before exposing agents to company systems.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews