# AI use cases for Long Horizon Harness (AMAP-ML/LongHorizon-Harness)

Repo: [AMAP-ML/LongHorizon-Harness](https://github.com/AMAP-ML/LongHorizon-Harness) | Use cases: 3 | Page: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/

LongHorizon-Harness is a loop that runs existing AI agents such as Claude Code, Codex, and OpenCode on desktop apps and the terminal for long tasks, checking results and saving verified progress along the way.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Run a long desktop task from the browser: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/run-long-desktop-task-from-browser/
- Delegate a terminal task to a chosen agent: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/delegate-terminal-task-to-named-agent/
- Split planning, execution, and auditing across models: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/split-roles-across-different-models/

## How to run a long desktop task from the browser with Long Horizon Harness

Repo: [AMAP-ML/LongHorizon-Harness](https://github.com/AMAP-ML/LongHorizon-Harness) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README

Start a multi-step computer task in a web dashboard and keep it running with checkpoints.

**Who it is for:** A person with a multi-step workflow across desktop apps who wants to start it, watch it, and answer approvals without using the terminal.

**The problem:** A single agent session often loses track of the goal or stops partway through a long workflow that spans several apps.

**The outcome:** A task that runs in rounds, with each round planned, executed, verified, and checkpointed, and a final plain-language reply.

### Steps

1. Complete the Installation steps in the README: install the harness and make sure one agent runtime is on PATH.
2. Run `lh-harness web` to start the Dashboard.
3. In the browser, start a new task, describe the goal, and choose a backend and model for the Manager, Executor, and Auditor.
4. Answer any approval requests and send extra instructions mid-run if the task needs changes.
5. Read the final reply and check the verified state before you rely on the result.

### Prompt to paste into your agent

```text
Use the LongHorizon-Harness repo at <repo path>. Start the Dashboard with lh-harness web, create a task for <goal>, and let it run until it reports completion. Expected output: the final reply and the verified state for <goal>. Verify that every round the Auditor accepted is listed as a checkpoint before you report success.
```

### Caveat

The README is truncated, so exact install commands and the full Dashboard workflow are not fully shown here; check the Installation and Dashboard sections in the repo.

Risk, credentials: The agent acts inside desktop apps, so any logged-in accounts it can reach may be used, and approvals should be reviewed before you grant them.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/run-long-desktop-task-from-browser/

## How to delegate a terminal task to a chosen agent with Long Horizon Harness

Repo: [AMAP-ML/LongHorizon-Harness](https://github.com/AMAP-ML/LongHorizon-Harness) | Difficulty: Advanced | Time: 1 hour | Basis: From the README

Send a command-line job to a specific agent backend and collect normalized results.

**Who it is for:** A developer who wants a long terminal job, such as code changes or a debugging session, run by a particular agent CLI with its own permissions.

**The problem:** Running a long terminal job by hand means watching it closely, and agent CLIs differ in how they take prompts and set permissions.

**The outcome:** A terminal task run by the chosen agent with role-scoped permissions, followed by a normalized result you can review.

### Steps

1. Install the harness and confirm the agent CLI you want, such as opencode, is available on PATH.
2. Run `opencode run prompt` to confirm the OpenCode CLI works on your machine before you use it through the harness.
3. Start a CLI task with the harness and select the agent with `--agent opencode`; check the CLI setup section for the exact run command. (inferred, check the README)
4. Review the normalized result and the Auditor's checks in the output before you accept the work.

### Prompt to paste into your agent

```text
Use the LongHorizon-Harness repo at <repo path> to run a terminal task for <goal> with the --agent opencode option. Expected output: the normalized result file or summary for <goal>. Verify that the tests or commands named in the goal were actually executed and passed.
```

### Caveat

The exact harness run command for CLI tasks is not visible in the truncated README, so the step marked inferred must be checked against the CLI setup section.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/delegate-terminal-task-to-named-agent/

## How to split planning, execution, and auditing across models with Long Horizon Harness

Repo: [AMAP-ML/LongHorizon-Harness](https://github.com/AMAP-ML/LongHorizon-Harness) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README

Assign a different model or backend to the Manager, Executor, and Auditor of one task.

**Who it is for:** A team that wants to balance quality, speed, and cost on a long task and does not want one model doing all three jobs.

**The problem:** One model handling planning, acting, and checking can be costly, and the same model may approve its own work without a real check.

**The outcome:** A task where each loop role runs on the model or backend chosen for it, with the Auditor checking the actual files and logs.

### Steps

1. Complete the Installation steps and confirm at least two agent runtimes are on PATH.
2. Run `lh-harness web` and open the Dashboard.
3. For each role, pick the backend and model: Manager for planning, Executor for actions, and Auditor for checks.
4. Start the task and watch which round the Auditor accepts or rejects.
5. Compare the final verified state with the goal to confirm the result.

### Prompt to paste into your agent

```text
Use the LongHorizon-Harness repo at <repo path> to run a task for <goal> with a different model assigned to the Manager, Executor, and Auditor. Expected output: a verified result and a list of accepted checkpoints. Verify that the Auditor inspected the real files or logs rather than relying on the Executor's claim.
```

### Caveat

Per-role model and backend selection is shown in the README for the Web workbench, but the full list of supported models per backend is not visible in the truncated text.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/amap-ml-longhorizon-harness/split-roles-across-different-models/