LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

LongHorizon-Harness, an open-source loop engineering system for AI agents, reached #1 on the Hugging Face Daily Papers weekly ranking for 2026-W32 and released v0.1.6 on 2026-08-15, adding OpenCode CLI support. The system, which works with Claude Code, Codex, OpenCode, and DeepSeek Harness, enables agents to autonomously complete real-world tasks across desktop apps and terminal CLIs over dozens of hours by planning, acting, verifying, checkpointing, and recovering in a continuous loop. It does not train new models but provides a durable execution loop around existing agents, with a browser-based dashboard and support for multiple backends.

Give Claude Code, Codex, OpenCode, or DeepSeek Harness a goal once. Keep it working across desktop apps and the terminal for dozens of hours. Plan → act → verify → checkpoint or recover → repeat — until the work is actually done. Usage one-command-full-visibility · The Loop loop-engineering-for-real-computer-environments · Computer Use desktop-apps-and-cli-one-continuous-task · Results hundreds-of-real-tasks-measured-gains · Project Website https://lh-harness.pages.dev · 简体中文 /AMAP-ML/LongHorizon-Harness/blob/main/README.zh-CN.md The model determines what an agent can do in one round. LongHorizon-Harness engineers the loop around it: what to do next, how to verify the result in the real computer, what progress to preserve, and how to continue after failure or context refresh. A Loop Engineering system for Claude Code, Codex, OpenCode, and DeepSeek Harness. One-command install, ready to run. LongHorizon-Harness turns existing agents into long-running computer-use systems. Across desktop apps and the terminal CLI, it continuously recovers the goal and verified state, selects the next bounded step, executes it with a fresh context, checks the actual result, and then checkpoints accepted progress or feeds failure evidence into the next round. It does not train a new model or replace an existing agent; it provides the durable execution loop around one. v0.1.6 · 2026-08-15 Added OpenCode https://github.com/anomalyco/opencode CLI support. LongHorizon-Harness can now run opencode run prompt as --agent opencode , with role-scoped read/write permissions, OpenCode API endpoint overrides, normalized JSON results, and CLI/config/doctor integration. The Web workbench can select OpenCode Harness and its model independently for each role. v0.1.5 · 2026-08-14 Added phase-1 DeepSeek Harness https://github.com/deepseek-ai/deepseek-harness CLI support. LongHorizon-Harness can now run dsh --profile headless as --agent deepseek harness , with an isolated DSH HOME , role-scoped read/write permissions, DeepSeek API endpoint overrides, normalized JSONL results, and CLI/config/doctor integration. The Web workbench can select DeepSeek Harness and its model independently for each role. GUI computer-use and MCP support will follow in a later phase; see the CLI setup 5-or-run-a-task-from-the-command-line . v0.1.4 · 2026-08-11 The new Dashboard has landed: a React/FastAPI workbench you can drive entirely from the browser. Start a task, choose a backend and model per role, answer approvals, send an instruction mid-run, and stop or restart a run. Launch it with lh-harness web ; see Run a task in the browser 4-run-a-task-in-the-browser-recommended . 2026-08-10 Added the Terminal-Bench 2.1 evaluation. v0.1.3 · 2026-08-07 Every run now ends with a plain-language reply that answers your task from the verified state alone. Tasks act on the directory you launched from by default, and the console reports each round as it happens. 2026-08-06 LongHorizon-Harness reaches 1 on the Hugging Face Daily Papers weekly ranking https://huggingface.co/papers/week/2026-W32 . v0.1.2 · 2026-08-06 Adds unified computer-use plugin management, stronger auditor read-only checks and role isolation, reliable process cleanup, and expanded doctor diagnostics. See Manage computer-use plugins manage-computer-use-plugins . 🚀 We’re iterating rapidly. Stay tuned promotional video 1440p.mp4 Open the promotional video 1440p MP4 Give LongHorizon-Harness an outcome. It repeatedly turns the remaining work into a bounded step, performs that step on the right computer surface, checks what actually happened, and carries the verified result into the next round. php flowchart LR S "Original goal +<br/ verified state" -- P "Plan the next<br/ bounded step" P -- A "Act in a desktop app or CLI<br/ with fresh context" A -- V "Verify files, UI, logs, and tests<br/ in the real environment" V -- |Pass| C "Checkpoint<br/ verified progress" V -- |Fail| R "Record evidence<br/ and recover" C -- D{"Task complete?"} R -- S D -- |No| S D -- |Yes| F "Verified result" This is Loop Engineering : designing the execution, verification, correction, and recovery loop around the agent — not just the prompt for a single turn. The roles are implementation boundaries inside the loop, not three agents independently growing their own versions of the task. | Loop responsibility | Role | What it owns | |---|---|---| 🧭 State and next step | Manager | Rebuilds each round from the original goal, verified progress, failure evidence, and remaining work | ⚡ Action | Executor | Starts with a fresh context and completes one clearly defined step in a desktop app or the CLI | 🔍 Ground truth | Auditor | Independently inspects the actual files, interfaces, logs, and tests instead of trusting the Executor's claim | Only results that pass independent verification become trusted task state. A rejected result remains evidence, not progress. When a context is refreshed, an action fails, or a deliverable does not pass inspection, the next round starts from the original goal and the last verified checkpoint, then continues from what remains. LongHorizon-Harness supports both GUI and CLI workflows. | 🖥️ Operate the desktop | ⌨️ Work in the terminal | |---|---| | 🌐 Click, type, scroll, and browse | 💻 Write and modify code | | 📊 Operate spreadsheets | | | 📄 Edit documents | 📦 Install dependencies and environments | | 🎨 Use design software | 🔧 Configure and debug systems | | 🧊 Operate 3D tools | 📁 Process files and data | One task can begin in a browser, move to the command line for data processing, continue in desktop software to produce an artifact, and return to the terminal for validation or debugging. The goal, progress, and evidence remain under the same state-management system throughout. LongHorizon-Harness is not tied to a specific model or agent backend. Existing models and agents connect through configuration without changing their original workflows. | Layer | Supported choices | | |---|---|---| | 🧠 | Models | Claude, GPT, Qwen, and other models exposed by an agent backend | | 🤖 | Agent backends | Claude Code, Codex CLI, OpenCode, DeepSeek Harness dsh , CLI-only in phase 1 , and custom AgentAdapter implementations | | 🎛️ | Role assignment | The Manager, Executor, and Auditor can each use a different model or backend | | 🖥️ | Execution environments | Local, with a pluggable Environment protocol | A lightweight AgentAdapter preserves each agent's native execution loop while LongHorizon-Harness coordinates role boundaries, verified task state, and cross-round progress around it. Use one model for all three roles, or combine different models and backends to balance quality, speed, and cost. LongHorizon-Harness is not demonstrated only on a handful of carefully selected success cases. We ran it on hundreds of complex tasks across GUI, CLI, and mixed computer environments: | Task domain | What the tasks involve | |---|---| 🌐 Web Frontend | Developing, fixing, and validating websites and web applications through browser interaction, developer tools, and code changes | 📊 Data Analysis & Visualization | Processing data, producing charts and dashboards, and checking analytical results and visual deliverables | 🛠️ Operations & Debugging | Investigating logs, networks, performance, and service failures; configuring, diagnosing, and repairing systems | 🎨 Design & Image Processing | Editing visual assets, matching design references, processing images, and verifying final visual quality | 🎮 Games & Interaction | Building, operating, and debugging games or interactive applications; checking interaction logic and runtime behavior | 📄 Documents & Presentations | Editing documents and slide decks, including content, formatting, references, layout, and final delivery | 🧊 Spatial Reasoning | Completing tasks involving spatial relationships, geometry, precise placement, and 3D operations | 🖥️ Desktop & System Settings | Operating desktop applications, files, and system settings across multi-application workflows | 🔬 Research & Education | Completing literature research, coursework, teaching materials, forms, and research-support workflows | 🎬 Creative Production | Producing presentations, video, audio, and other media while coordinating assets across tools | ⚙️ Engineering & Computing | Using CAD, EDA, scientific software, development tools, and cloud or DevOps toolchains | 🎫 Personal Services | Handling event ticketing, everyday services, games, and visual-search workflows | 🏛️ Administration & Compliance | Completing office, legal, policy-sensitive form, institutional, and safety-aware submission workflows | 💼 Business & Finance | Handling market analysis, procurement, loans, sales, reimbursements, and cross-application enterprise workflows | 🏥 Healthcare | Completing medical quality-control, insurance, immunization, and structured health-form workflows | GUI + CLI completionWeaveBench | Full desktop-task completionOSWorld 2.0 | Code + CLI successTerminal-Bench 2.1 · 24% fewer tokens | | Benchmark | Metric | Claude Code | LongHorizon-Harness | Gain | |---|---|---|---|---| WeaveBench 114 tasks | PassRate | 51.8 | 80.7 | +28.9 | WeaveBench | Overall | 0.702 | 0.835 | +0.133 | OSWorld 2.0 108 tasks | Binary | 2.8 | 8.3 | 3.0× | OSWorld 2.0 | Partial | 21.5 | 35.2 | +13.7 | Terminal-Bench 2.1 | Success rate | 69.7 | 77.2 | +7.5 | All rows use Qwen 3.7-Plus as the backbone and Claude Code as the execution backend. Full result tables and case trajectories are available on the LongHorizon-Harness project website https://lh-harness.pages.dev . Steps 1–2 are once per machine; step 3 is once per project. Then run tasks from the browser step 4 or the command line step 5 . | Needed for | | |---|---| | uv tool install brings its own; a pip install uses yours. PATH : , https://github.com/openai/codex installing-and-running-codex-cli codex , https://docs.anthropic.com/en/docs/claude-code/getting-started claude , or https://github.com/anomalyco/opencode opencode dsh Node.js https://nodejs.org 20 or later ^22.19.0 or =24.0.0 . Platform status:Currently tested on macOS. Windows support is included but has not yet been thoroughly tested. Run lh-harness doctor at any point to check all of the above; see Verify the environment verify-the-environment . uv tool install lh-harness or: pip install lh-harness Upgrade later with uv tool upgrade lh-harness or pip install --upgrade lh-harness . Skip this if your tasks never touch the GUI. Otherwise install the one that matches your agent. No plugin is enabled by default, and one install covers every project on the machine. Using Codex: lh-harness plugin install codex-computer-use Using Claude Code, or both agents: lh-harness plugin install open-computer-use codex-computer-use is the official plugin bundled with the Codex CLI and only works with Codex. open-computer-use is distributed on npm, needs Node.js 20+, and drives both agents. Both need OS permissions that must be granted by hand on macOS . See Manage computer-use plugins manage-computer-use-plugins for that, for clawdcursor as a third option, and for how each one is wired. cd /path/to/your/project lh-harness init This creates ./.lh-harness/config.toml without replacing an existing file; use lh-harness init --force to regenerate. Open it and adjust the defaults. Every field is documented in Configuration reference configuration-reference . lh-harness web --workspace-root . This opens the workbench at http://127.0.0.1:8799/ . Everything happens there: start a task, pick a backend and model per role, answer approval requests, send an instruction mid-run, and stop or restart a run. --workspace-root sets the default working directory for tasks created there; the remaining options are listed under Dashboard commands dashboard-commands . TASK="Inspect the current directory and summarize its files." lh-harness run --task "${TASK}" --agent codex Explicit CLI arguments such as --agent override the matching values in ./.lh-harness/config.toml for that run; drop them to use the configured defaults. To use the phase-1 DeepSeek Harness CLI backend, install its official npm package, provide a DeepSeek API key, and select deepseek harness : npm install -g @deepseek-ai/dsh If your npm mirror has not synced the package: npm install -g @deepseek-ai/dsh --registry=https://registry.npmjs.org dsh --version export DEEPSEEK API KEY="sk-..." Optional for a private or compatible endpoint: export DEEPSEEK BASE URL="https://your-endpoint.example.com" lh-harness doctor lh-harness run --task @task.md --agent deepseek harness \ --model deepseek-v4-flash --no-dashboard To make DeepSeek Harness the project default, put this in ./.lh-harness/config.toml : run agent = "deepseek harness" model = "deepseek-v4-flash" dashboard = false Then use LongHorizon-Harness as usual: lh-harness run --task @task.md The LongHorizon Web workbench also exposes DeepSeek Harness CLI in each role's Harness selector and offers deepseek-v4-flash plus a custom model ID. Export the provider environment variables before starting the Web server so its worker processes inherit them: export DEEPSEEK API KEY="sk-..." export DEEPSEEK BASE URL="https://your-endpoint.example.com" lh-harness web --workspace-root . The adapter runs dsh --profile headless , gives every run an isolated DSH HOME , uses workspace-write for executors, and uses read-only for the Manager and auditors. --api-key maps to DEEPSEEK API KEY , --base-url maps to DEEPSEEK BASE URL , and LH HARNESS DSH BINARY can select a non- PATH binary. DeepSeek Harness is still a developer preview; this phase intentionally does not expose its Web UI, computer-use plugins, MCP config, or --mcp-add-dir . Its headless profile currently returns only the final answer, so intermediate DeepSeek tool events are not streamed into the trajectory; the upstream positional task interface also means the task text is visible in the child process argument list while an episode is running. The agents work in the directory you launched from, so the task acts on your real project. Set workspace or --workspace to point somewhere else. ./.lh-harness/ itself stays off limits, so the run's own logs and state are never mistaken for task content. The Dashboard opens in your browser automatically, and the console prints one line per role as the run progresses. At the end you get a plain-language reply that answers your request from the verified state alone, and says so plainly if the task did not finish. Every run is stored under ./.lh-harness/runs/<run-id / ; the full report, including that reply, stays in the run's logs/report.json . lh-harness doctor doctor is read-only. It reports the Python runtime, the agent CLIs, Node.js, and plugin state, and exits non-zero when a required check fails. Agent CLIs are verified by running <binary --version , not just by finding them on PATH , so one that is present but broken is reported as a failure instead of OK. This catches the Windows case where a Microsoft Store desktop install leaves a zero-byte codex.exe alias on PATH that is not the CLI; doctor prints how to fix it. It also checks PyPI https://pypi.org/project/lh-harness for a newer version. To check on its own: lh-harness check-update lh-harness run reads ./.lh-harness/config.toml automatically. Precedence is: - Explicit CLI arguments - Values in ./.lh-harness/config.toml - Built-in defaults Task text, run IDs, and API keys are deliberately not configurable here; they stay command-line or environment inputs so they never land in a file you might commit. | Field | Default | Description | |---|---|---| agent | "codex" | Backend for every role unless a role overrides it: codex , claude code , opencode , or deepseek harness . | model | "gpt-5.6-sol" | Model for every role unless a role overrides it. Must be a model the chosen backend exposes. | env | "local" | Execution environment. Only local today. | runs root | "./.lh-harness/runs" | Where run directories are created. Each run gets <runs root /<run-id / . | workspace | commented out | Working directory the agents operate in. Defaults to the directory lh-harness was started from, so a task acts on your real project; set it to isolate the run somewhere else. | harness dir | commented out | Where harness task state is written. Defaults to the run's own harness/ , keeping it out of the workspace. | log dir | commented out | Where logs are written. Defaults to the run's own logs/ . | base url | commented out | OpenAI-compatible endpoint override, for a proxy or a self-hosted model. | prompt language | "en" | Language of the harness-generated prompts and reports: en or zh . Does not restrict the task language. | claude mcp config | commented out | Path to a .mcp.json for Claude Code. Overrides the installed plugin. | codex mcp config | commented out | Path to a mcp servers. TOML for Codex. Overrides the installed plugin. | mcp add dirs | | Extra directories the MCP server may read. Claude Code rejects these, because its role isolation requires task files to live inside the workspace. | max rounds | 30 | Upper bound on Manage-Execute-Audit rounds before the run stops. | dashboard | true | Start the web dashboard with each run. | dashboard port | 0 | Dashboard port; 0 lets the OS pick a free one. | Per-episode limits in seconds. One episode is a single role invocation, not the whole run. | Field | Default | Description | |---|---|---| manager | 600 | Planning the next step. | gui executor | 1800 | Executing a GUI/visual subtask. | cli executor | 1800 | Executing a CLI/non-GUI subtask. | auditor | 600 | Verifying a subtask. Applies to both auditors. | Each role can take its own agent and model , so you can pay for a strong model only where it matters: a capable Manager and Auditor with a cheaper Executor, for example. Every field is commented out by default, meaning "inherit". Resolution walks the chain until it finds a value: gui executor → executor → run .agent / run .model cli auditor → auditor → run .agent / run .model | Section | Falls back to | Covers | |---|---|---| run.roles.manager | run | The scheduler role | run.roles.executor | run | Both executor roles | run.roles.gui executor | executor | GUI/visual subtasks | run.roles.cli executor | executor | CLI/non-GUI subtasks | run.roles.auditor | run | Both auditor roles | run.roles.gui auditor | auditor | GUI audit | run.roles.cli auditor | auditor | CLI audit | run.roles.final response | manager | The closing reply written for you | Every field above also has a CLI flag --agent , --max-rounds , --gui-executor-model , --auditor-timeout , and so on that overrides it for a single run. Run lh-harness run --help for the full list. If a Manager, Executor, or Auditor reaches its local episode timeout, the run keeps the partial trajectory and recorded task state, then lets the next Manager round inspect the real workspace and recover. The timeout remains an agent execution timeout; it is not treated as proof of a provider network failure. Repeated timed-out rounds still trigger the Dashboard's human-review gate. Computer-use setup is intentionally separate from task execution: doctor only reports status, and lh-harness run never installs, removes, or changes plugins. All changes go through lh-harness plugin . List the available plugins with their install state, supported agents, and homepages: lh-harness plugin list | Plugin | Source | Agents | Platforms | |---|---|---|---| codex-computer-use | Official plugin bundled with the Codex CLI | codex | whatever your Codex build offers | open-computer-use | npm | codex , claude code clawdcursor clawdcursor https://github.com/AmrDab/clawdcursor codex , claude code Installing needs no agent flag. Every agent the plugin supports is configured, since the per-agent difference is only one more config file: lh-harness plugin install clawdcursor One install covers every project on the machine. It installs the package, runs whatever consent or permission step the plugin needs on the current OS, and writes one MCP config per agent under ~/.lh-harness/plugins/ . Agents missing from PATH are skipped; --agent narrows the selection, and --no-activate skips the permission step on a headless machine. lh-harness run then loads the right server automatically. When several are installed, the first available one wins: codex-computer-use open-computer-use clawdcursor --claude-mcp-config and --codex-mcp-config override that choice. plugin list and doctor both print which plugin each agent will load and whether its permissions are granted. To remove one: lh-harness plugin uninstall clawdcursor GUI access stays scoped to the harness. The npm plugins live entirely inside ~/.lh-harness/ and are passed per run, so ~/.codex/config.toml , ~/.claude.json , and the user-scope MCP registries are never touched. codex-computer-use is the unavoidable exception: Codex loads it from its own registry, so codex plugin add records it there. codex-computer-use needs manual grants on macOS. It raises no permission dialog, so an unauthorized GUI call just fails. The install opens the two panes for you; tick Codex Computer Use under Privacy & Security → Accessibility and → Screen & System Audio Recording , then re-run the install to verify. On Windows there is nothing to grant, but the harness has to run in a signed-in desktop session and stay unelevated. Any missing prerequisite is printed during install. Any MCP server can be passed to the agents, not just computer-use ones. Each backend reads its own native format; nothing is translated between them. Claude Code takes a .mcp.json file through --claude-mcp-config : { "mcpServers": { "computer-use": { "command": "/path/to/mcp-server", "args": "--option", "value" , "env": { "EXAMPLE VARIABLE": "value" } } } } Codex takes a TOML file of mcp servers.<name tables through --codex-mcp-config , matching ~/.codex/config.toml : mcp servers.my-server command = "/path/to/mcp-server" args = "--option", "value" mcp servers.my-server.env EXAMPLE VARIABLE = "value" Pass the config for the backend in use, plus any directory the server needs to read: lh-harness run --task @task.md --agent codex \ --codex-mcp-config /path/to/mcp.toml \ --mcp-add-dir /path/to/mcp/files Both flags can be given together when roles use different backends, and --mcp-add-dir may be repeated. The equivalent environment variables are LH HARNESS CLAUDECODE MCP CONFIG , LH HARNESS CODEX MCP CONFIG , and LH HARNESS MCP ADD DIRS , the last separated by : on macOS/Linux and ; on Windows. Prefer letting the server read API keys from its environment over writing them into the config file. lh-harness run --task @task.md --dashboard Monitor a live run lh-harness dashboard Browse completed and active runs lh-harness web --workspace-root . Serve the workbench for another directory dashboard and web start the same workbench and accept the same options; web reads as the plain service entry point when the workbench is what you want, not a side effect of a run. | Option | Description | |---|---| --workspace-root | Default workspace for runs created from the workbench default: current directory | --runs-root | Base directory holding runs default: ./.lh-harness/runs | --log-dir | Pin one run's log directory instead of browsing --runs-root | --host / --port | Bind address default: 127.0.0.1:8799 ; --port 0 lets the OS pick | --auth-token | Bearer token, required for any non-loopback --host also LH HARNESS WEB TOKEN | --no-open | Do not open the URL in a browser | | Option | Description | |---|---| --task | Task text or @task.md | --agent | claude code , codex , opencode , or deepseek harness CLI-only in phase 1 | --env | local | --max-rounds | Maximum number of Manage-Execute-Audit rounds; the CLI default is 30 | --dashboard | Start live monitoring and human intervention | --no-dashboard | Disable a Dashboard enabled by the project configuration | Run a longer task from a file and open the Dashboard: lh-harness run --task @task.md --dashboard The Dashboard shows every round's plan, execution result, audit evidence, and reason for rework. It also provides human gates when a task completes, becomes blocked, needs input, or fails repeatedly. | 📋 Plan | ⚡ Execution | 🔍 Audit | ♻️ Rework | |---|---|---|---| | What happens next | What the agent did | What the environment proves | Why another round is needed | Every run is stored in an isolated runs/<run-id / directory. The complete task state and audit trail make the agent's progress inspectable, recoverable, and reproducible. | Run record | What it preserves | |---|---| 📋 Task state | Original goal, requirements, verified progress, and remaining work | 🧾 Event stream | What happened throughout the run | 🔍 Audit reports | Evidence and acceptance decisions for every round | 🧠 Role trajectories | Manager, Executor, and Auditor inputs and outputs | 📁 Workspace | Files and artifacts produced during execution | ✅ Final report | The verified outcome of the task | eval/ provides frozen reproduction suites for three benchmarks: | Directory | Benchmark | Description | |---|---|---| eval/WeaveBench-harness/ | eval/OSWorldv2-harness/ eval/TB-harness/ See each directory's README.md or README.zh-CN.md for environment setup, parameters, and launch commands. The nested Harness / cua harness code is a frozen compatibility copy used for evaluation; new integrations should use src/lh harness/ . @article{longhorizonharness2026, title={LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks}, author={Ziyu Ma and Hailang Huang and Shun Zou and Yong Wang and Shidong Yang and Yiming Hu and Fei Wei and XiangXiang Chu}, journal={arXiv preprint arXiv:2608.01964}, year = {2026}, url = {https://arxiv.org/abs/2608.01964} } Operate the whole computer. Preserve verified progress. Keep working until the task is done.