AG-2 in Practice #6 – Human-in-the-Loop (HITL) Workflows

#ai #ag2 #architecture

Welcome back!

In the last post, we built a custom pattern with conditional logic and tool use. But in many real-world apps, you don’t want to leave everything to agents. You want a human to step in before critical actions — especially in tasks involving judgment, risk, or sensitive data.

That’s where Human-in-the-Loop (HITL) comes in.

Why HITL?

Human-in-the-loop is useful when:

You want approval or intervention before agents execute a plan
You want to review intermediate results
You’re building co-creative tools (e.g., coding assistants, design AI)
You need accountability (e.g., legal, medical, or finance use cases)

Step 1: Add a Human Agent

AG-2 allows you to include a human as an agent in the loop. You can interact via terminal, notebook, or integrate a UI later.

Here’s how to set it up:

from ag2 import Agent, HumanAgent, Pattern, Conversation
import os

os.environ["OPENAI_API_KEY"] = "your-api-key-here"

# AI Agent
writer = Agent(
    name="writer",
    llm="openai/gpt-4",
    system_message="You write summaries based on user prompts.",
)

# Human Reviewer
human_reviewer = HumanAgent(
    name="human_reviewer",
    system_message="You are a human reviewer. Approve or suggest edits.",
)

Step 2: Create a Review Flow with a Human

from ag2 import Orchestrator

orchestrator = Orchestrator(
    agents=[writer, human_reviewer],
    rules=[
        {"from": "user", "to": "writer"},
        {"from": "writer", "to": "human_reviewer"},
        {"from": "human_reviewer", "to": "user"},
    ]
)

conv = Conversation(orchestrator=orchestrator)
conv.send("Write a summary of the history of the internet.")

Now the human will be prompted to review the AI’s output before it’s finalized.

Step 3: Try It in Terminal

When run from your terminal, AG-2 will pause and ask for human input at the human_reviewer step:

python hitl_review.py

Example flow:

Writer generates a summary
Terminal pauses and shows output to the human
You (the human) type: “Looks good, approve.”
Flow ends or loops based on what you say

Step 4: Add HITL to Complex Patterns

HITL can be used inside custom patterns, as a QA reviewer, escalation point, or fallback agent when the system gets stuck.

You can mix and match AI + humans freely — even simulate multi-role humans if needed.

What’s Next?

In the next post, we’ll explore:

The AG2 Studio for visual workflows
Inspecting logs, replays, and debugging sessions
Deploying and iterating on agent systems without writing code

Keep coding

DEV Community