OpenHands¶

简体中文

Local editor/terminal events and real code execution in an isolated working directory; not the complete OpenHands platform.

Paper information¶

Field	Details
Paper	OpenHands: An Open Platform for AI Software Developers as Generalist Agents
Institution	All-Hands-AI / Carnegie Mellon University and collaborators.
First public release	2024-07-23
Author code	All-Hands-AI/OpenHands.
Local adapter or method key	`openhands`
Local reproduction code	`src/auto_research/agent_research/`

Paper method and evidence¶

OpenHands is an open software agent platform. It unifies terminal, editor and browser actions in an observable event stream and isolates execution in a sandbox for bug fixes, implementation and repository maintenance.

flowchart LR
  A[Agent action] --> S[Editor or terminal sandbox]
  S --> E[Observable event stream]
  E --> N[Next action]

Original paper figure¶

The original figure shows the agent's event-driven platform. From the original paper; copyright remains with the authors.

Core equation¶

\[ e_{t+1}=\operatorname{Sandbox}(a_t),\qquad a_{t+1}=\pi(g,e_{\le t+1}). \]

Reported results¶

The paper evaluates the platform and multiple controllers on software-engineering benchmarks including SWE-bench. No production A/B.

Local reproduction¶

Local editor and terminal actions write to the event stream. The terminal runs a fixed regression test in a separate temporary directory and reports actions, edit hashes, command output and success.

auto-research agent-eval --method openhands --benchmark swebench-local \
  --episodes 12 --seed 42

Stable metrics. Browser actions currently retain interface semantics only. No live browser, full controller, or official SWE-bench container.