OpenHands: An Open Platform for AI Software Developers as Generalist Agents

By mid-2024 many teams were building agents that write code, run commands and browse the web, inspired by Cognition's Devin. Each had its own way of representing actions, executing them, keeping history and sandboxing. That made agents hard to compare and hard to build on. Xingyao Wang, Graham Neubig and a large group of collaborators from universities and companies built OpenHands (formerly OpenDevin) as a shared platform.

It has three core pieces. An agent is a class with a step method: it receives the current state and returns one action. The state includes an event stream, an ordered log of every action and observation, from the agent, the user and the environment, plus bookkeeping such as accumulated model cost. The runtime executes actions inside a Docker container per session: a bash shell, a Jupyter IPython kernel for Python, and a Playwright-driven Chromium browser using BrowserGym's action set. The agent never touches the host; the runtime turns each action into an observation and appends it to the stream.

Around these sit a library of reusable tools (AgentSkills, Python functions preloaded in the sandbox), a delegation action that lets one agent hand a subtask to another, support for running on any user-provided Docker image, a chat interface that shows what the agent is doing, and an evaluation harness covering 15 benchmarks. The default generalist, CodeActAgent, acts mainly by writing code. With one interface, it came close to specialist systems on software engineering, trailed them on web navigation, and beat prior results on graduate-level science questions.

Whether you adopt OpenHands or build your own, its architecture is a good reference for the plumbing every customer agent needs. Most production problems with agents are plumbing problems: where code runs, what was done, and how to debug it.

Keep a complete, ordered event log. Every action and observation, with who issued it. It is your audit trail, your debugging tool and your evaluation data, and it lets you replay or resume a session.

Separate decision from execution. The agent proposes actions; a runtime executes them under its own rules. That is where you enforce sandboxing, permissions and resource limits, independent of which model or agent design you use.

Sandbox per session. A fresh container with only the customer's workspace mounted limits the blast radius of a bad command or an injected instruction.

Make agents swappable. A common interface lets you test a new model or agent design against the same tasks without touching the runtime.

Delegate to specialists. A generalist that hands web tasks to a browsing agent, or reviews to a checking agent, keeps each component simpler.