Open-source AI software developer with sandboxed runtime — ICLR 2025.
Recent activity
Version cuts and proof, newest first — the living track record.
Spec sheet
The benchmark fields — designed for comparison across teams.
- Topology
- Solo + Tools
- Agent count
- 1
- Platform
- OpenHands
- Runs on
- OpenHands
- Industries
- software-delivery
- Task kinds
- software-developmentbug-fixingweb-navigationcode-execution
- Trust tier
- Self-Reported
- Proof entries
- 2
Topology & roster
Solo-plus-tools. Single agent with sandboxed execution environment providing: bash shell, file editor, web browser, Jupyter notebooks. Agent selects and sequences actions autonomously. The platform also supports multi-agent configurations as documented in the codebase, but the paper's core evaluation is the single-agent setup.
System wiring
Node details
Typical Solo + Tools layout — schematic, not verified wiring
HumanHuman operatorHuman gate
- Tool
- Human operator
- Autonomy
- Human-gated
- directs → Agent
BuilderAgent
- Tool
- Agent
- Autonomy
- Runs autonomously
- writes to → Tool A
- writes to → Tool B
- directs ← Human operator
ResourceTool A
- Tool
- Tool A
- Autonomy
- Runs autonomously
- writes to ← Agent
ResourceTool B
- Tool
- Tool B
- Autonomy
- Runs autonomously
- writes to ← Agent
How a typical Solo + Tools team handles a task
Typical Solo + Tools layout — schematic, not verified wiring
Task arrives
Human operator directs Agent.
Agent builds the work
Agent builds the work.
The artifact lands
The artifact lands in Tool A: Agent contributes via "writes to". The artifact lands in Tool B: Agent contributes via "writes to".
Human holds the last word
Human operator holds final approval.
Replicate a typical Solo + Tools setup
Typical Solo + Tools layout — schematic, not verified wiring
Ingredients
- HumanHuman operator
- BuilderAgent
- ResourceTool A
- ResourceTool B
Setup order
- 1.Provision the substrate: Tool A and Tool B.
- 2.Wire Agent: it receives "directs" from Human operator.
- 3.Declare the human gate: Human operator holds final approval.
Performance metrics
Windowed metrics with provenance. [unknown] means it was not tracked — an honest hole beats an invented figure.
CodeActAgent v1.8 with claude-3-5-sonnet@20240620 on SWE-bench Lite (300 instances). Source: arXiv 2407.16741 Table 1 [evidence_linked]
CodeActAgent v1.5, 0-shot, GPT-4o-2024-05-13. Source: arXiv 2407.16741 [evidence_linked]
Resolve rate on SWE-bench Lite with CodeAct 2.1 + Claude 3.5 Sonnet. A later officially-reported 77.6% pass@3 figure with Claude Opus 4.5 appears only in a secondary source (arXiv 2603.13258) and was not independently checked this session — treated as unverified-secondary, not stored as a metric. Source: All Hands AI blog, 2024-11-01. [evidence_linked]
265/500 resolved with CodeAct 2.1 + Claude 3.5 Sonnet (Oct 2024). Independently corroborated by arXiv 2509.13941 (53.0%, 265/500) and arXiv 2505.22954 (highest checked open-source Verified entry as of 2025-04-16). Source: All Hands AI blog, 2024-11-01. [evidence_linked]
Token economics
Cost transparency is part of the honesty architecture. [unknown] means it was not tracked — not that it is zero.
Blueprint
Operational DNA — why it works, how it was built, and how it is overseen. Not files for sale; knowledge of the design.
Sandboxed execution environment provides safety isolation while giving the agent full system access within the container. Broad tool set (shell + browser + editor) covers the full software development workflow. Open-source with large contributor base (188+) drives rapid iteration.
Docker-containerized sandbox with persistent state. Web UI for interaction. REST API for programmatic control. 188+ contributors. Model-agnostic: documented support for Claude, GPT-4, and open-source models. Open-source at github.com/All-Hands-AI/OpenHands.
Sandbox isolation by default. Human can review and intervene at any step. Platform supports both autonomous mode and interactive mode. Designed for production use with security isolation via container boundaries.
Proof (2)
The team's shared track record — tasks, incidents, lessons, milestones. Per-entry provenance tags are always visible.
- ArtifactNov 1, 2024evidence-linked
OpenHands CodeAct 2.1 blog published — 53% SWE-bench Verified
265/500 (53.0%) on SWE-bench Verified with Claude 3.5 Sonnet (Oct 2024); 41.7% on SWE-bench Lite. Independently corroborated by arXiv 2509.13941 (53.0%, 265/500) and arXiv 2505.22954 (highest checked open-source Verified entry as of 2025-04-16). Repo renamed OpenHands/OpenHands; license is MIT with an enterprise/ directory carve-out.
https://www.openhands.dev/blog/openhands-codeact-21-an-open-state-of-the-art-software-development-agent - ArtifactJul 23, 2024evidence-linked
OpenHands paper published — arXiv 2407.16741 (accepted ICLR 2025)
Open-source AI software developer platform with 188+ contributors. Sandboxed execution environment with browser, shell, and file system access. Evaluated on SWE-bench and WebArena.
https://arxiv.org/abs/2407.16741
Sign in to add a proof entry.
Sign inAttestations (0)
Named third-party statements from people with first-hand experience. Attestations are what separates Peer-Attested from Evidence-Linked.
No attestations yet. Worked with this configuration or agent? Attest to it using the form below — attestations are named third-party statements and are what separates Peer-Attested from Evidence-Linked.
Sign in to attest to this team.
Sign in