OpenHands

Sources#

• Coding Agents and Technical Debt
• EVOMAL: Self-Poisoning in Self-Evolving Coding Agents

Summary#

OpenHands is an open-source coding-agent platform, and also the company that maintains and sells it. It began as OpenDevin — Wang et al., "OpenHands: An Open Platform for AI Software Developers as Generalist Agents", arXiv 2407.16741, ICLR 2025 — and is cited under that paper in several of this wiki's research sources. Unlike Claude Code and Codex, its source is public, which is why it doubles as the corpus's default third-party agent scaffold: when a paper needs a harness it doesn't own to run a model against SWE-Bench or Terminal-Bench, it usually reaches for OpenHands.

The four-repo platform#

Its own July 2026 engineering post (Coding Agents and Technical Debt) is the corpus's most detailed public accounting of a coding-agent codebase, precisely because the pieces are separate repos:

Totals: ~1.05M lines, 5,679 merged PRs, 1,778 of them bug fixes (31%) in the twelve months to 2026-07-08, with 179 contributors on the app repo alone. The runtime/surface split is the same shape Shared Harness, Differentiated Surfaces describes at OpenAI, here shipped as four artifacts instead of one.

Commercial position. OpenHands sells the maintained agent, and its published argument is that you should rent the runtime rather than fork it — customize via prompts, MCP servers, skills/plugins, or the SDK, in that order. See Harness Build-vs-Buy, where that argument is developed with its evidence caveats.

Where it shows up in this corpus#

Mostly as somebody else's experimental apparatus, which is a useful independent signal about its standing:

• Single-Rollout Optimization — SAO's SWE-Bench Verified results are run in the OpenHands scaffold (Qwen3-30B-A3B backbone, 300 turns, 128k context)
• Orchestration-Plan Simulation — one of OrchBench's four real-execution frameworks for cross-framework validation, alongside Claude Code, SWE-mini, and Crush
• Knowledge-Centric Self-Improvement — a reported Terminal-Bench 2 comparator (13.9% on the Haiku 4.5 entries table)
• Agent Self-Poisoning (the CREATE-Path) — one of three scaffolds EvoMal (arXiv 2608.25776, empirical) reran its skill-library self-poisoning attack on unmodified, running its default CodeActAgent against DeepSeek-V4-Pro through a litellm proxy with an identical precomputed retrieval cache. Its generic-attacker rate is 22.2% against mini-SWE-agent's 41.8%, but task-family targeting brings it to 66.7% on pytest — the same ceiling as the research scaffold — and the counter-prompt takes it to 0.0%
• Also cited as a baseline or agent-vendor label in raw/swe-pruner-pro and raw/ai-code-review-practitioner-discourse

Connections#

• Harness Build-vs-Buy — OpenHands' own build-vs-buy argument and the twelve-month GitHub data behind it; the page carries the vendor-interest caveats
• Shared Harness, Differentiated Surfaces — its app/SDK/Canvas/CLI split is the runtime-plus-surfaces architecture shipped as separate repos
• Codex — the OpenAI harness it benchmarks itself against (7,688 merged PRs, ~1.32M lines)
• Hermes Agent — the other open agent in the same comparison (7,736 merged PRs, ~1.75M lines)
• Claude Code — the closed-source peer, absent from OpenHands' comparison for that reason
• Orchestration-Plan Simulation — used as one of four real harnesses OrchBench validates its simulator against
• Single-Rollout Optimization — the scaffold SAO's coding-agent RL results are produced in
• Knowledge-Centric Self-Improvement — a Terminal-Bench 2 comparator system
• Agent Self-Poisoning (the CREATE-Path) — the CREATE-path attack reproduces inside its default CodeActAgent unmodified, which is part of the paper's evidence that self-poisoning is a property of the retrieve-author-persist loop rather than of one research harness

Sources#

• Coding Agents and Technical Debt — Rajiv Shah (OpenHands), 2026-07-28 (case-study): the four-repo breakdown, twelve-month activity data, and the fork/build-vs-buy argument
• Referenced as a research scaffold or baseline in: Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning, OrchBench: Evaluating Multi-Agent Orchestration Plans in Isolation via Deterministic Simulation, Knowledge-Centric Self-Improvement, SWE-Pruner Pro: The Coder LLM Already Knows What to Prune, 3100 Opinions on Code Review in an AI World: Building Causal Theory from Practitioner Discourse
• EVOMAL: Self-Poisoning in Self-Evolving Coding Agents — Wu, Shi, Q. Li, Zhao, X. Li, Adams, Hassan & Ni (Queen's University), arXiv 2608.25776, 2026-08-26 (empirical): §7 and App. C.3 (Figure 5) — OpenHands as one of two production scaffolds the attack was rerun on, with per-family CREATE-path rates and the counter-prompt result