1
# Human In The Loop
2
3
The failure mode of an approval gate is not that it lets something bad through. It is
4
that it asks so often, and so vaguely, that the human starts approving without
5
reading. This harness defines what genuinely needs a human, how to ask so the answer
6
is informed, and how far a given approval extends.
7
8
## When to use
9
10
- Agents that can write, send, deploy, delete, or spend.
11
- Unattended or scheduled runs, where a bad action is found long after the fact.
12
- Any workflow where a mistake is expensive to undo.
13
14
## Install
15
16
```bash
17
npx openagents-cli add openagents/human-in-the-loop
18
```
19
20
| Runtime | Installed to |
21
|---|---|
22
| `claude-code` | `.claude/skills/human-in-the-loop/` |
23
| `cursor` | `.cursor/rules/human-in-the-loop/` |
24
| `codex` | `.codex/skills/human-in-the-loop/` |
25
| `openai-agents` | `agents/human-in-the-loop/` |
26
| `langgraph` | `graphs/human-in-the-loop/` |
27
| `generic` | `.openagents/human-in-the-loop/` |
28
29
## What is in the package
30
31
- `HARNESS.md` - when to stop, how to ask, and the scope of an approval.
32
- `gates.md` - a classification of actions by reversibility, with the gate each needs.
33
34
## License
35
36
MIT
37