OpenHands

Sweet spot: a developer or small team that wants to experiment with fully autonomous coding agents and is comfortable with Docker and a bit of setup. OpenHands is the most hackable credible option — you can swap models, modify agent logic, and plug it into your existing workflow through the GitHub App. As a research platform, it is arguably the most interesting agent codebase in open source. Failure modes. Autonomy is still brittle. Even with GPT-4o or Claude Sonnet 4, complex multi-file refactors fail often enough that human review is non-negotiable. Cost per run is higher than interactive IDE use because autonomous agents loop more. And for day-to-day developer productivity, a supervised copilot (Cursor, Claude Code) will get you more done per dollar. What to pilot. Pick three well-scoped bugs from your backlog — each should have a clear test and a bounded blast radius. Run OpenHands on all three via the GitHub App. Measure success rate, time spent reviewing, and total cost. If at least one passed cleanly and the other two needed only minor fixes, OpenHands has earned a slot in your automation stack; if none made it past review, the economics do not work yet for your codebase.

OpenHands is an open-source autonomous coding agent that can plan, write, run, debug, and iterate on code across a full development environment. It runs inside a sandboxed Docker container with file-system access, a web browser, and a terminal — giving it the same surface area a human developer uses. Originally launched as OpenDevin in response to Cognition's Devin demo, it was renamed to OpenHands in 2024 and evolved into one of the most active open-source agent projects. The agent architecture separates Planning from Execution, supports multiple agent implementations (CodeActAgent, BrowsingAgent, VisualBrowsingAgent), and integrates with every major LLM via LiteLLM. Interfaces include a web UI (the default), a CLI, a headless mode, and a GitHub App that responds to issues and comments. OpenHands Cloud offers a hosted version with included credits for teams that do not want to manage Docker containers. It is MIT-licensed, benchmarks near the top on SWE-Bench among open agents, and has a rapidly growing contributor base. A primary use case is as a starting point for custom autonomous engineering agents rather than a polished end-user product.