TL;DR — All Hands AI shipped OpenHands 1.0 on September 8, the production release of the autonomous coding agent formerly known as OpenDevin. Paired with Qwen3-Coder-480B it resolves 68% of SWE-bench Verified tasks; with Claude Sonnet 4.5 and extended thinking, 72%. The headline is not the benchmark but the hardening: a production-grade Docker sandbox that finally makes self-hosted autonomous coding defensible.
Introduction
Autonomous coding agents have a security problem that has kept most teams on the sidelines. To be useful, an agent must execute the code it writes — running tests, installing packages, starting services — and doing that on a shared host without isolation is a non-starter. OpenHands 1.0 is the open-source answer: an agent that keeps your code and API keys inside your own infrastructure, with the sandbox story that production demands (Source : ByteIota — OpenHands 1.0: Self-Hosted Coding Agent With Safety Sandbox).
The Benchmark
The headline result puts OpenHands at or above commercial rivals on the most widely cited evaluation for autonomous software engineering. Paired with Qwen3-Coder-480B at up to 100 turns, it resolves 68% of SWE-bench Verified tasks. Swapping in Claude Sonnet 4.5 with extended thinking lifts that to 72% (Source : TechPillow — OpenHands 1.0 Ships as Open-Source AI Coding Agent).
For context, Devin has claimed around 77.8% on its own SWE-1.7 benchmark, but the comparison is imperfect across different model pairings and turn budgets. What matters is that an open-source agent with a bring-your-own-model architecture is now within striking distance of the leading commercial product, at a fraction of the infrastructure cost.
The Sandbox Is the Real Story
The 1.0 release is less about new capabilities than about making existing ones production-safe. Every task OpenHands executes now runs inside an isolated Docker container with configurable resource limits — CPU, memory, and network access — enforced per agent session. Execution runs as non-root via a SANDBOX_USER_ID of 1000, closing the most obvious privilege-escalation path (Source : ByteIota — OpenHands 1.0: Self-Hosted Coding Agent With Safety Sandbox).
The most distinctive feature is a built-in LLM-based security analyzer that rates every action LOW, MEDIUM, or HIGH before execution. That is a meaningful departure from static policy alone: instead of only allow-listing commands, the agent uses a second model pass to judge risk contextually. It is not bulletproof, but it converts the security question from a binary "sandboxed or not" into a graduated, inspectable signal.
Why Self-Hosting Just Got Viable
All Hands AI, a San Francisco company founded in 2024, maintains OpenHands under a permissive license, with no gate on the underlying agent framework and no publicly disclosed external funding (Source : TechPillow — OpenHands 1.0 Ships as Open-Source AI Coding Agent). The project began as OpenDevin, a community experiment in applying model reasoning to real GitHub issues rather than isolated code-completion tasks.
Version 1.0 marks the transition from research prototype to something an engineering team can deploy. The bring-your-own-model design means the benchmark number is not the ceiling — teams can plug in the strongest model they can afford, on infrastructure they control. In a market where most autonomous coding agents are closed and cloud-hosted, that is the differentiator.
FAQ
What is OpenHands?
The production release of OpenDevin, an open-source autonomous coding agent that files and resolves GitHub issues end-to-end.
How does it compare to Devin?
It resolves 68–72% of SWE-bench Verified depending on the model, competitive with commercial rivals at a fraction of the cost.
Is it safe to run?
Every task runs in an isolated Docker container with non-root execution, resource limits, and an LLM-based security analyzer that rates each action.
What models does it support?
Bring-your-own-model. The 1.0 benchmarks used Qwen3-Coder-480B and Claude Sonnet 4.5.
Further Reading
- ByteIota — OpenHands 1.0: Self-Hosted Coding Agent With Safety Sandbox
- TechPillow — OpenHands 1.0 Ships as Open-Source AI Coding Agent
— The Agent Report
Cet article a été initialement publié sur The Agent Report.
Top comments (0)