⚡ Workflows
Human-in-the-Loop
Embed mandatory human review gates at critical points in an otherwise autonomous workflow. Agents handle all routine work; humans are invoked only when decisions exceed defined risk thresholds, ensuring speed without sacrificing control.
Bundle files
📄ORCHESTRATION.md— Orchestration
1# ORCHESTRATION.md
2
3## Diagram
4
5 ┌──────────────────────────────────────────┐
6 │ Workflow Orchestrator │
7 └────────────────────┬─────────────────────┘
8 ▼
9 ┌──────────────────────────────────────────┐
10 │ Agent (autonomous work) │
11 └────────────────────┬─────────────────────┘
12 ▼
13 ┌─────────────┐
14 │ Checkpoint │
15 │ reached? │
16 └──────┬──────┘
17 no │ yes
18 ┌────────────────┴──────────────────┐
19 ▼ ▼
20 ┌──────────┐ ┌─────────────────────────┐
21 │ Continue │ │ Human Review Gate │
22 │ autonomy │ │ (approve/reject/edit) │
23 └──────────┘ └────────────┬────────────┘
24 approved │ rejected
25 ┌──────────────┴──────────────┐
26 ▼ ▼
27 ┌──────────────┐ ┌──────────────────┐
28 │ Continue │ │ Agent reworks │
29 │ workflow │ │ with feedback │
30 └──────────────┘ └──────────────────┘
31
32
33## Workflow
34Human-in-the-Loop
35
36## Objective
37Embed mandatory human review gates at critical points in an otherwise autonomous workflow. Agents handle all routine work; humans are invoked only when decisions exceed defined risk thresholds, ensuring speed without sacrificing control.
38
39## Roles
40- **Workflow Orchestrator:** manages the overall flow and knows which checkpoints require human approval.
41- **Executing Agents:** carry out autonomous work between checkpoints.
42- **Human Reviewer:** receives formatted review requests, approves/rejects/edits, and provides feedback.
43- **Gate Agent:** packages checkpoint outputs into a clear review brief and waits for human decision.
44
45## When to use
46- Any AI workflow where decisions have irreversible real-world consequences (publishing, payments, deployments).
47- Regulated industries requiring human sign-off on AI-generated outputs.
48- New workflows in their first N runs before trust is established.
49- When the agent's confidence falls below a defined threshold.
50- Gradual automation rollout: start with many gates, remove them as reliability is proven.
51
52## Checkpoint triggers
53A human gate is required when any of the following are true:
54- Action is irreversible (delete, publish, pay, deploy to production).
55- Estimated impact exceeds defined threshold (e.g. cost > $500, users affected > 1000).
56- Agent confidence score < configured minimum.
57- Action touches a sensitive category (PII, legal, compliance, security).
58- First time this action type has been performed.
59
60## Protocol
611. Orchestrator and agents execute work autonomously until a checkpoint trigger is met.
622. Gate Agent packages a review brief: what was done, what is proposed, impact analysis, confidence, and a clear approve/reject/edit choice.
633. Review brief is delivered to Human Reviewer via configured channel (Slack, email, UI).
644. Human Reviewer responds within timeout:
65 - **Approved:** Orchestrator continues.
66 - **Rejected with feedback:** Agent reworks and returns to the same gate.
67 - **Edited:** Human's edit is accepted as-is; Orchestrator continues with that version.
685. If timeout expires: escalate; workflow halts until response.
696. All gate decisions are logged with human identity, timestamp, and rationale.
70
71## Rules
72- Gates must be configured before the workflow starts — no ad-hoc gates.
73- Human must receive enough context to make an informed decision in under 2 minutes.
74- Agents must never retry a rejected action without incorporating human feedback.
75- Approval from the wrong role level is not valid (match reviewer to risk level).
76- Gate logs are immutable audit records.
77
78## Deliverables
79- Completed workflow output (post all approvals).
80- Gate decision audit log: reviewer, decision, timestamp, feedback, retries.
81- Automation coverage report: % of steps handled autonomously vs. gated.
82- Recommended gates to relax after N successful runs.
83
Lines: 83 | Words: 500
Install
Copy bundle to your OpenClaw workspace.
curl -fsSL https://raw.githubusercontent.com/cerealskill/openclaw-agents/main/install.sh | bash -s workflow human-in-the-loop ENRate this agent
Loading...
Sign in to rate this agent
Includes
- ✓ ORCHESTRATION.md
Info
- Slug
- human-in-the-loop
- Lines
- 83
- Words
- 500