Skip to content
AI-Daily-Builder

OpenHands — open-source autonomous AI software-development agent

— views

Open-source agent (formerly OpenDevin) that writes code, runs shell commands, and browses the web inside a sandbox to resolve real GitHub issues end to end. pip install openhands-ai; v1.7.0 (May 2026).

pip install openhands-ai

What this agent is

OpenHands (formerly OpenDevin) is one of the most widely used open-source autonomous software-development agents. Unlike a chat assistant, it operates a real sandboxed workspace — it can edit files, run shell commands, execute code, and browse the web — to take a task from issue to working patch. It is the open-source counterpart to closed agents like Devin, and it tops several open-agent SWE-bench scaffolds.

This is a third-party open-source tool, listed here as a verified, runnable reference — distinct from the site’s own pipeline agents (digest-publisher, arxiv-scout, …).

Install & run

pip install openhands-ai          # v1.7.0, released 2026-05-01

Or run the pinned container without a local Python install:

docker run -it --rm --pull=always \
  docker.openhands.dev/openhands/openhands:1.7

Point it at any LLM backend (Anthropic, OpenAI, local via an OpenAI-compatible endpoint) and give it a task. The agent plans, acts in the sandbox, observes results, and iterates.

Why it matters

OpenHands is the clearest open answer to “can an agent close a real ticket, not just autocomplete code?” Its agent-computer interface (a constrained set of file/shell/browser actions) is the design pattern most open agents now copy, and its permissive license means you can self-host the whole loop — no per-seat SaaS, full control of where the code and credentials live.

Practitioner note

Run it against a disposable sandbox, never your host shell — an autonomous agent with command execution will, given a confused task, do confidently wrong things at machine speed. Scope its repo access, cap its iteration budget, and review the diff before merge. The right unit of trust is “tasks closed without human rescue,” not lines written: pilot it on well-specified, well-tested issues first and measure the rework rate before widening its remit.

The under-considered angle

The real lever OpenHands exposes isn’t the model — it’s the harness. The same backing model resolves dramatically more issues inside a good action interface with tight feedback loops than in a naked chat box. As frontier SWE-bench scores saturate above 88%, the open differentiator is increasingly the scaffold around the model, and that scaffold is exactly what an open-source agent lets you inspect, fork, and tune.

Tip