OpenHands

OpenHands now separates its application from the OpenHands Software Agent SDK. The reviewed SDK repository owns agents, tools, conversations, workspaces, events, a Python SDK, an Agent Server with REST/WebSocket APIs, and a TypeScript client. The application consumes those contracts.

Evidence status: source-reviewed on 2026-09-23 at 5b36cac (v1.49.5). Agentic Arena has not run it locally.

• Software Agent SDK
• Official SDK architecture
• Agent Server

Distinguishing design

The reasoning-action loop is stateless over an event history: it prepares context, queries the model, validates proposed actions, executes tools, and appends observations. Conversation services own lifecycle and persistence. Condensers manage history, while a security analyzer and confirmation policy can inspect actions before execution.

A caller can embed the Python SDK or operate a long-running Agent Server. The server exposes conversations, workspace files and commands, and streaming events. Workspaces can be local or ephemeral remote environments, including container and cluster-backed configurations. This makes the execution provider an explicit comparison input.

What Agentic Arena should test

Use the Python SDK first for a minimal run, then repeat through Agent Server to measure boundary cost and recovery. Capture events independently and verify that pause/resume, confirmation, process loss, and workspace cleanup preserve the same task semantics. Treat local and remote workspace runs as separate configurations.

OpenHands teaches us to separate an embeddable agent kernel, a remotely operable workspace service, and user-facing clients. It also shows why an event history needs explicit ownership: a stateless loop is recoverable only if the conversation and workspace records agree.