Harness board · hand-maintained
/
Last checked 26 Sep 2026
19 harnesses indexed
The board · maintained by hand
AI Harness Leaderboard
What runs the loop. The programs that hold a model, its tools and its permissions together for long enough to finish a job. There is no composite score here: harness quality depends on the job, so this board compares what each tool is, how it isolates work and who it is for. Tap any column to re-sort.
⚙ How to read this board
Order is editorial. Open any row for its full spec sheet.
Not a score. Judgement calls on what each harness is for, checked against vendor docs.
| # | Harness | Interface | License | Isolation | Best for |
|---|---|---|---|---|---|
| 1 | Claude Codecoding AnthropicTerminal, desktop, web and IDE. The reference implementation of a coding agent with real permissions. |
Terminal, desktop, web, IDE | Proprietary (Anthropic) | Permissioned tools, opt-in sandbox | A coding agent with real permissions |
| 2 | Claude Agent SDKlibrary AnthropicClaude Code's loop, tools and context management as a library. Python and TypeScript only — other languages drive the CLI as a subprocess. |
Imported library | Proprietary (Anthropic) | Inherits the host process | Building custom agents on Claude Code's loop |
| 3 | Managed Agentshosted AnthropicHosted REST API for long-running agents. Anthropic runs the agent and the sandbox, so you don't operate session infrastructure. |
REST API | Proprietary, hosted | Anthropic-run sandbox | Long-running agents without running infra |
| 4 | Codex CLIcoding OpenAISandboxed by default and open source. The usual pick when those two are hard requirements. |
Terminal | Open source | Sandboxed by default | Open-source + sandbox as hard requirements |
| 5 | Gemini CLIcoding GoogleApache-2.0, strong on whole-repository context. The free tier ended 18 Jun 2026, so it's now a budgeted decision. |
Terminal | Apache-2.0 | Configurable | Whole-repository context (now paid) |
| 6 | Editor agentsin-editor Cursor, Zed, Copilot, othersAgent loops living inside the editor. Tighter feedback, narrower blast radius. |
IDE integration | Mixed / proprietary | Editor workspace | Tight feedback with a narrow blast radius |
| 7 | Hermes Agentagent open source · self-hostedAn always-on agent for small businesses: it watches a mailbox, form or queue and acts on what lands. Covered in our setup and small-business guides. |
Self-hosted service | Open source | Your own host | Always-on mailbox / queue automation |
| 8 | OpenClawcoding open sourceOpen-source, model-agnostic terminal coding agent — Claude Code's ergonomics without the lock-in, MIT-licensed. |
Terminal | MIT | Permissioned tools | Claude Code ergonomics without lock-in |
| 9 | Ampcoding SourcegraphAgentic coding built for big codebases, with deep repository search sitting behind the loop. |
CLI + editor | Proprietary (Sourcegraph) | Workspace | Large codebases with deep search |
| 10 | OpenCodecoding open sourceProvider-agnostic terminal agent: bring any model, keep the same loop and permissions. |
Terminal | Open source | Permissioned tools | One loop across any model |
| 11 | Prime Agentcoding Prime IntellectSelf-improving harness: one persistent Python kernel as the only tool |
Terminal + ACP | MIT | Background subagent sessions | Long-horizon self-improving work |
| 12 | DeepSeek Harnesscoding DeepSeekEverything is a plugin — including the agent loop |
Terminal | MIT | OS-level sandbox + session log | Assembling a harness from parts |
| 13 | OpenHandscoding All-Hands AIFormerly OpenDevin: self-hosted, sandboxed, model-agnostic |
Terminal + web | MIT | Optional Docker sandbox | Self-hosted automation you can audit |
| 14 | Gooseagent Agentic AI FoundationBlock's agent, now under Linux Foundation governance |
Desktop + CLI + API | Apache-2.0 | Local runtime, MCP extensions | Provider-agnostic work on your machine |
| 15 | Qwen Codecoding QwenLM (Alibaba)Gemini CLI lineage, tuned for Qwen models |
Terminal | Apache-2.0 | Configurable | Running Qwen models in a strong loop |
| 16 | Aidercoding Aider-AIGit-native pair programmer: one commit per edit |
Terminal | Apache-2.0 | Git-tracked edits | Short supervised runs with clean history |
| 17 | Picoding earendil-worksMinimal harness: a system prompt under a thousand tokens |
Terminal | Open source | Permissioned tools | A minimal harness over a strong model |
| 18 | Grok Buildcoding xAIxAI's harness: memory that survives between sessions |
Terminal + web + mobile | Harness open-sourced | Hosted service + CLI | High-volume coding on a subscription |
| 19 | GitHub Copilot CLIcoding GitHubGitHub-native terminal agent with auto model routing |
Terminal | Proprietary | Permissioned tools + GitHub context | Teams living in GitHub and Copilot seats |
Editorial index · hand-maintained, not a feed
Hand-maintained, not a feed. Checked 26 Sep 2026 against Anthropic's model docs, OpenAI's model list and Google's Gemini docs. Vendor facts come from those pages; the read on each harness is judgement, not vendor copy.