Paper summary: OpenHands: An Open Platform for AI Software Developers as Generalist Agents

Problem

LLM agents that can act, writing code, running commands and browsing the web, were developing quickly, but mostly in separate codebases that were hard to compare or build on.

Method

OpenHands (formerly OpenDevin) is an open platform where agents interact with the world the way a human developer does: writing code, using a command line and browsing the web. It provides:

• a way to implement new agents
• sandboxed environments for safe code execution
• coordination between multiple agents
• built-in evaluation benchmarks

The paper’s Figure 1 shows the user interface, where you can watch files, commands and the agent’s browser.

Key results

The authors evaluate agents on 15 challenging tasks, including software engineering (SWE-bench) and web browsing (WebArena). The project is a community effort released under a permissive licence.

Limitations

The paper’s future-work section names current weaknesses:

• multimodal support (images, video, spreadsheets) isn’t yet principled
• agents still struggle with complex tasks
• agents perform badly when editing long files