OpenHands

OpenHands (formerly OpenDevin) is the leading open-source autonomous software engineering agent — achieving 72% on SWE-bench Verified (resolving real GitHub issues) and shipping a production-ready SDK in late 2025. With 64K+ GitHub stars and an MIT licence, it is the top open-source alternative to commercial agents like Devin.

Architecture Deep Dive → OpenHands Architecture Breakdown — event-sourced SDK design, CodeAct execution paradigm, Docker sandbox isolation, and multi-agent orchestration patterns.

Why It's in Trial

OpenHands crossed from research project to production platform in 2025. The key shift was the V1 SDK release — a modular, stateless, event-sourced architecture designed for running thousands of agents concurrently in the cloud rather than just locally.

What makes it stand out:

• SWE-bench Verified: 72% resolution rate using Claude Sonnet 4.5 with extended thinking — top-tier performance on the industry's standard benchmark for autonomous coding agents
• Model agnostic: Bring any LLM — Claude, GPT-5-Codex, Gemini, or local models via Ollama
• Sandboxed execution: Agents run inside Docker containers — file writes, shell commands, and web access are isolated from your host system
• Web UI + GitHub integration: Resolve GitHub issues directly from the UI; no terminal required
• AMD partnership: Optimised for running on AMD Radeon workstations with open-source local models — a strong option for air-gapped/on-premises deployments
• MIT licence: Full auditability, no proprietary components in the core

The Production SDK (V1)

The V1 SDK is composable: four packages — SDK, Tools, Workspace, Server — with clear separation. The event-sourced state model means a crashed agent can be resumed from any point in its execution history. This matters at scale: when running 100 parallel agents, individual failures shouldn't require manual recovery.

Compared to Devin

When OpenHands Struggles

• Large monorepos with complex build systems. OpenHands clones your repo into a Docker container and works from there. For repos with multi-minute build times or complex dependency graphs (Bazel, Gradle multi-project), the agent spends significant time and tokens on build failures that a human would avoid by knowing the local environment. Claude Code's "work in your environment" approach handles this better because it inherits your local toolchain.
• Tasks requiring deep domain context. The 72% SWE-bench score means 28% of real GitHub issues are not resolved. The failures tend to cluster on tasks requiring understanding of project-specific conventions, implicit architectural constraints, or multi-file coordination across distant parts of the codebase. If your issues routinely require reading 10+ files to understand context, expect lower success rates.
• Ops overhead of self-hosting. "Free" is misleading — you need to provision Docker infrastructure, manage API keys, monitor agent runs, and handle the inevitable cases where an agent gets stuck in a loop burning tokens. For small teams without DevOps capacity, Devin's managed SaaS may actually be cheaper when you account for ops time.

For Engineering Managers

OpenHands is the answer to "can we get Devin-level capabilities without the $500/seat cost?" For teams willing to host it on their own infrastructure and bring their own LLM API keys, the performance and functionality are comparable at a fraction of the price. But factor in the ops cost of self-hosting — if your team doesn't have Docker infrastructure expertise, the "free" label is deceptive.

Getting Started

docker pull ghcr.io/all-hands-ai/openhands:main

docker run -it --rm \
  -e LLM_API_KEY=your-anthropic-key \
  -e LLM_MODEL=claude-sonnet-4-6 \
  -p 3000:3000 \
  ghcr.io/all-hands-ai/openhands:main

Open http://localhost:3000 and start assigning tasks.

Key Characteristics

Further Reading

• OpenHands Architecture Breakdown — full deep dive
• OpenHands V1 SDK announcement
• SWE-bench Verified leaderboard