OpenHands vs SWE-agent
The two most cited open-source agentic coding frameworks. SWE-agent (Princeton, 2024) introduced the Agent-Computer Interface idea. OpenHands (AllHands, formerly OpenDevin) went production: Docker sandboxes, a browser tool, a growing plugin ecosystem.
Verdict
OpenHands is the production path: 40k+ GitHub stars, active weekly releases, Docker sandboxing, and the best SWE-Bench Verified score of any open scaffold (~66% with Sonnet 4.5, ~77% with Opus 4.5 internal harness). SWE-agent is the research path: minimal, text-first, and the canonical reference implementation for agent-computer interfaces. If you are writing a paper, use SWE-agent. If you are shipping a product, use OpenHands.
Benchmark snapshot
pass@1, best reported
Open scaffolds on SWE-Bench Verified — same models, different scaffolds
OpenHands ships more tools (browser, jupyter, bash, file editor); SWE-agent ships fewer, more targeted tools. The mini-SWE-agent 100-LOC descendant is often used for apples-to-apples model comparison.
Architecture side-by-side
OpenHands exposes a richer toolset wrapped in Docker. SWE-agent exposes a minimal tool interface tuned for LLM ergonomics.
Architecture
OpenHands
Docker runtime, browser, event memory
Architecture
SWE-agent
Minimal Agent-Computer Interface
Radar
Scaffold capability profile (0-10)
At a glance
| Attribute | OpenHands | SWE-agent |
|---|---|---|
| Org | AllHands (ex-OpenDevin) | Princeton NLP |
| Paper | CodeAct / OpenDevin 2024 | SWE-agent NeurIPS 2024 |
| License | MIT | MIT |
| Runtime | Docker sandbox | Local filesystem / container |
| Tools | bash, file_edit, browse, jupyter, custom MCP | ls, find_file, open, scroll, edit, submit |
| Edit format | str_replace + full-file rewrite | line-range linear edit |
| Memory | Event-history + condenser | Full trajectory in context |
| Best Verified (Opus 4.5) | 77.6% | 72.0% |
| Stars | ~44k | ~14k |
| Best for | Production self-hosted agents | Research baselines, papers |