How OpenHands 1.0 Brings Production-Ready Sandboxing to Open-Source Coding Agents

OpenHands has reached its 1.0 release, adding production-ready Docker sandboxing, built-in security policies and a plugin system, with the open-source autonomous coding agent now completing roughly 68% of SWE-bench Verified tasks when paired with a frontier model.

Sep 8, 2026
How OpenHands 1.0 Brings Production-Ready Sandboxing to Open-Source Coding Agents
Source: Youtube - Openhands

The open-source coding agent world just got its "production-ready" stamp. OpenHands, the autonomous software engineering agent formerly known as OpenDevin, has shipped its 1.0 release, adding production-grade Docker sandboxing, built-in security policies, resource limits, and a plugin system — the kind of guardrails that previously gave companies an excuse to stick with commercial alternatives instead of self-hosting.

The headline number is performance: OpenHands paired with a frontier model completes roughly 68% of SWE-bench Verified tasks, a 500-issue benchmark drawn from real GitHub repositories. That figure puts it in the same range as commercial agents charging per task. A few details show how the release scales across budgets:

  • With open-weight Devstral 24B as the backend, OpenHands scores around 46.8% — matching what funded commercial product Devin 2.0 has publicly reported
  • Self-hosted runs cost roughly $0.20 to $1.05 per resolved task on H100 infrastructure, depending on model and task complexity
  • Every agent action — file edits, shell commands, test runs — now executes inside an isolated Docker container, keeping code and credentials off the controller host

One developer testing the release put it plainly: "the harness does not care" which model sits behind it, meaning teams can route routine work through cheaper open models and save frontier models for harder architectural problems.

According to a roundup of the day's agent news, the release lands the same week GitHub expanded Copilot Workspace to run multiple coordinating agents — a sign that both the open-source and commercial sides of autonomous coding are converging on the same idea: agents are only as useful as the sandbox and safety rails built around them.