What it does
OpenHands provides an interface for software agents to interact with a workspace through tools such as shell commands and file operations. Runtime and model integrations are configured separately.
Why it’s interesting
The action-observation loop and runtime boundaries are useful architecture topics. Treat generated code as untrusted and review every action before using it in a real repository.
Real-world use case
Evaluate an agent on a disposable sample repository and compare its proposed patch with a human-written fix.
What I can learn
- 01
Agent action-observation loops
Study the agent-computer interaction loop: the agent chooses a terminal, file-editing, or browser action, receives its result, then uses that observation to plan the next step.
- 02
Runtime and workspace isolation
Learn how isolated runtime containers and workspace boundaries let an agent run commands and modify project files without affecting unrelated host data.
- 03
Coding-agent evaluation workflows
Explore how coding-agent evaluations score real issue-solving tasks by checking reproduction steps, code changes, and whether relevant tests pass.
One thing to explore
Runtime and agent implementations
Follow the openhands/runtime implementation and connect it to the concepts above.
Explore the source (opens in a new tab)10-Minute Challenge
Inspect a sample agent run
Before you start: A browser and source access. Avoid mounting the host Docker socket or real project files for this reading exercise.
- Read the current security and local setup documentation.
- Review how actions are represented and returned as observations.
- Trace one shell action through the runtime implementation.