3.3 Module 3 · Single-Agent Design Patterns

Approval Gates & Human-in-the-Loop

Design approval workflows by setting confidence thresholds and risk categories. Then simulate 20 requests to see which get auto-approved, flagged for review, or escalated.

Approval Flow Builder Request Simulation Engine

Approval Flow Builder

Configure your approval gates. Set confidence thresholds for auto-approval, define risk categories, and choose escalation paths for different scenarios.

Confidence Thresholds
Auto-Approve above threshold
85%
Flag for Review between thresholds
50%
Escalate below review threshold
Anything below 50% confidence is automatically escalated to a human supervisor.
Risk Category Overrides

Some actions should always require approval regardless of confidence. Toggle categories that must always be reviewed.

Approval Flow
Incoming Request
Check Risk Category
High-risk category
Always escalate
Normal category
Check confidence
Confidence ≥ 85%
Auto-Approve
50% - 84%
Flag for Review
Confidence < 50%
Escalate

Key insight: The best approval systems are not binary (approve/reject). They use graduated confidence thresholds combined with category-based overrides. This lets low-risk routine tasks flow freely while catching edge cases and high-stakes decisions for human review.

One more gate you don’t control: frontier models now ship with their own misalignment monitors that can pause or stop a long run without asking you. OpenAI says to expect exactly that with GPT-6 Astra at launch — in ChatGPT and Codex you’re asked to review and continue, in the API the task just stops. Design for it: checkpoint long runs so a pause costs you minutes, not the whole job.

Request Simulation Engine

Run 20 simulated requests through your approval logic. Watch in real time as each request is classified based on your thresholds and risk categories.

Approved: 0
Review: 0
Escalated: 0

Configure your approval flow above, then click "Run Simulation" to process 20 sample requests.