Human-in-the-Loop Systems

Build hybrid systems where agents automate routine work and humans handle complex decisions

Designing Approval Workflows

When agents can't fully automate a task, they don't just dump it on humans. Instead, they do the work and present recommendations for human approval. This reduces human cognitive load while maintaining oversight.

The Three-Stage Approval Pattern

1️⃣
Agent Analyzes

Agent gathers context, evaluates options, formulates a recommendation with reasoning

2️⃣
Human Reviews

Human examines the recommendation, checks reasoning, makes the final call

3️⃣
Agent Executes

If approved, agent carries out the action; if rejected, agent adapts or escalates further

What Makes a Good Approval Request?

Actionable

Clear, specific proposal. "Refund $450 to account #12345" not "Customer is unhappy"

Contextualized

All relevant info included. Order history, customer status, policy rules—no need to dig around

Reasoned

Agent explains why. "Policy XYZ applies because..." helps humans verify logic quickly

Binary Decision

Approve or reject, not open-ended. Minimizes decision fatigue—yes/no is faster than "what should we do?"

Interactive: Approval Dashboard Simulator

You're a human supervisor. The agent has analyzed these requests and is asking for your approval. Click "Let Agent Propose" to see the agent's recommendation, then approve or reject:

💰
Refund $450
Customer reported damaged item, photo evidence provided
⏸️Pending
🗑️
Delete Account
User requested account deletion via support ticket
⏸️Pending
⬆️
Upgrade Plan
Business customer wants to upgrade from Pro to Enterprise
⏸️Pending
💡
Key Insight

Notice how the agent doesn't just say "there's a problem"—it analyzes the situation, proposes a solution, and explains its reasoning. This transforms your role from problem-solver to quality-checker, dramatically reducing cognitive load.

Previous