OpenHands
Vintage: 2026-08. Official SDK + architecture docs as synthesized in 2026-08-27-best-agent-harnesses-for-programming. GitHub star count is a thin fetch (README body not extracted). SWE-bench.org / OpenHands Index were not fetched.
OpenHands
One-line summary: MIT, model-agnostic Software Agent SDK (All Hands / OpenHands) — the Python/REST engine behind OpenHands CLI and Cloud. A harness substrate, not an IDE recap.
What it is
A coding-specific agent SDK with bash/file/browser/MCP tools, a Docker/K8s server, and an OpenAI-compatible endpoint. Official architecture splits openhands.sdk (agent loop, tools, skills, condenser, security), openhands.tools, openhands.workspace (local / Docker / remote), and openhands.agent_server. Same agent code swaps LocalWorkspace → DockerWorkspace → RemoteAPIWorkspace.
The loop is a ReAct-style reasoning-action loop inside Agent + Conversation. Production path is container/remote isolation, not Codex-style local Seatbelt/seccomp policy language.
Why it matters to this thread
It is the leading open, model-agnostic harness this source surveyed — the counterpart to vendor-locked claude-code / Claude Agent SDK and codex-cli. Docs claim OpenHands is “becoming the preferred harness for evaluating LLMs on coding tasks.” Official Terminal-Bench 2.0 does not make it the TB winner. Leave both claims; they may refer to different benches and dates.
Key facts (from 2026-08-27-best-agent-harnesses-for-programming)
- License: MIT. Model-agnostic.
- GitHub: 85,281 stars (thin fetch; README body not extracted).
- Terminal-Bench 2.0: OpenHands + Claude Opus 4.5 at 51.9%±2.9 — below Codex CLI + GPT-5.5 (82.2%) and several closed CLIs.
- Workspace swap is the load-bearing architecture claim: same agent code, different isolation.
Strengths (from this source)
- Open SDK; same agent code across local / Docker / remote workspaces.
- Coding-specific tool set (bash/file/browser/MCP) without a single-lab model lock.
Weaknesses / open claims
- TB 2.0 score is not competitive with the top vendor CLI or the eval-specialized harnesses. See terminal-bench-harness-swing.
- “Preferred eval harness” + SOTA on SWE-bench / SWT-bench / multi-SWE-bench is an OpenHands docs claim. This pass did not fetch SWE-bench.org or the OpenHands Index. Do not treat as confirmed.
Open questions
- Do the SWE-bench / SWT-bench / multi-SWE-bench claims hold on the current public tables?
- Is OpenHands a daily-coding harness or primarily an eval/container harness?