human-in-the-loop

name	human-in-the-loop
description	Patterns for incorporating human judgment into agent workflows — confirmation gates, escalation triggers, approval workflows, and graceful degradation when human input is unavailable.
compatibility	["VS Code","Cursor","Windsurf","Claude Code"]
metadata	{"category":"Uncategorized","tags":["uncategorized"],"maturity":"beta","audience":["developers"]}
allowed-tools	["bash","git","grep","find"]

Human-in-the-Loop

Use this skill when an agent can keep working autonomously on low-risk tasks but needs explicit human judgment before taking actions that are destructive, ambiguous, high-stakes, novel, or below its confidence threshold.

When to Involve Humans

• Destructive operations: delete, deploy to production, modify infrastructure
• Ambiguous requirements: multiple valid interpretations exist
• High-stakes decisions: security changes, data migrations, public-facing changes
• Novel situations: no prior pattern or precedent in the knowledge base
• Confidence below threshold: agent isn't sure the approach is correct

Confirmation Gate Pattern

Before executing a gated action:

• Present a concise plan summary
• Describe the expected impact
• Include a rollback strategy
• Wait for explicit approval instead of auto-proceeding on timeout
• Log the human's decision for an audit trail
• If no human is available, queue the action rather than skipping it silently

Escalation Triggers

Escalate when any of the following occurs:

• Error count exceeds threshold (3+ failures on the same task)
• Token budget is approaching its limit while work remains incomplete
• Conflicting requirements are detected
• A security-sensitive file modification is detected
• A test suite regression is introduced

Approval Workflows

Synchronous

The agent pauses, asks the user for approval, and waits for a response before continuing the gated step.

Asynchronous

The agent creates a PR or issue for review, continues other independent work, and checks back later for a decision.

Delegated

The agent asks a more capable agent to review or validate the approach, then falls back to a human if that delegation fails.

Graceful Degradation

When human input is unavailable:

• Document what is blocked and why
• Complete what can be done without approval
• Create a clear handoff for when the human returns
• Never silently skip a confirmation gate
• Set a timeout with an explicit action deferred status

Anti-Patterns

• Auto-approving after timeout because it defeats the purpose of the gate
• Asking for confirmation on every trivial action, which creates approval fatigue
• Asking for binary yes/no decisions without context, which prevents informed judgment
• Blocking all work when one step needs approval instead of parallelizing the rest