Phase 6: Dify & Visual AI App Builders · 50 min · Dify · OpenAI API · Anthropic API

Dify Agent Configuration — Tools, Strategies, and Autonomy

Dify agents: visual configuration, production deployment.

Hiring signal: Dify's agent app type is the fastest way to deploy a production AI agent. Being able to configure agent strategies (Function Calling vs ReAct), select tools, set iteration limits, and connect knowledge bases demonstrates the full-stack AI building skill that Dify-specific roles require.

What you will learn

Configure Dify agents: select agent strategy (Function Calling or ReAct), choose LLM model, set iteration limits
Add tools to Dify agents: built-in tools (search, calculator, DALL-E), custom tools (HTTP API, workflow), and knowledge bases
Design agent system prompts for specific business use cases
Test and debug Dify agents using the agent preview and trace view

The Problem

You've built Dify chatbots and workflows. Now you need an AI agent — an autonomous assistant that can use tools to complete tasks. In LangChain, this requires 100+ lines of Python defining tools, agent executor, and error handling. In Dify, it's a visual configuration: select agent strategy, choose tools, write a system prompt, set iteration limits.

Dify's agent app type is the fastest way to deploy a production AI agent. But configuring it correctly — choosing the right strategy, selecting the right tools, writing an effective system prompt, and setting safety limits — determines whether the agent works or fails.

Dify agents trade code for configuration
Instead of writing Python tool definitions and agent executors, you configure tools visually in Dify's UI. The trade-off: less flexibility than LangChain, but dramatically faster deployment. For most business use cases, Dify's agent configuration is sufficient and 10x faster to production.

The Concept

Dify Agent Strategies

Strategy	How It Works	Best For	Cost
Function Calling	LLM selects a function and provides arguments — one step	Known tool selection, structured tasks	Low (1 LLM call)
ReAct	LLM iterates: think → call tool → observe → repeat	Dynamic tasks, multiple tool calls needed	Higher (multiple LLM calls)

FUNCTION CALLING AGENT:
  User: "Create an order for Chicago to Dallas, $2125, J.B. Hunt"
  LLM: selects create_order(origin="Chicago", dest="Dallas", rate=2125, carrier="J.B. Hunt")
  → Done. One LLM call, one tool call.

REACT AGENT:
  User: "Create an order for Chicago to Dallas, J.B. Hunt"
  LLM: "I need the rate. Let me look it up."
  → calls lookup_rate(Chicago, Dallas) → $2125
  LLM: "I need to verify the carrier. Let me search."
  → calls search_carrier("J.B. Hunt") → MC# 242962
  LLM: "I have all info. Creating order."
  → calls create_order(...) → ORD-123456
  → Done. Three LLM calls, three tool calls.

A user asks "What's the weather in Chicago and should I delay my shipment?" The agent needs to: check weather, assess if it affects shipping, and potentially recommend delay. Which strategy is best?

ReAct is best because the task requires multiple steps with reasoning: (1) call weather API, (2) observe the result and reason about whether it affects shipping, (3) potentially call a shipment API to check status, (4) decide on recommendation. Function Calling only does one step — it can't iterate based on observations. The ReAct agent adapts: if weather is fine, it says "no delay needed"; if weather is severe, it recommends delay and checks alternatives.

Dify Agent Tools

Tool Type	Examples	Configuration
Built-in	Google Search, Calculator, DALL-E, Wikipedia	Toggle on/off
Custom HTTP	Any REST API	Define URL, method, parameters, auth
Workflow	Another Dify workflow as a tool	Select published workflow
Knowledge Base	RAG retrieval	Select connected knowledge base

Agent Configuration Options

Setting	What It Controls	Recommended
Strategy	Function Calling or ReAct	FC for simple, ReAct for complex
Model	Which LLM powers the agent	GPT-4o-mini (cost) or GPT-4o (quality)
Tools	Which tools the agent can use	3–5 tools max
Iteration limit	Max ReAct cycles	5–10
System prompt	Agent's role and constraints	Specific to use case
Memory	Conversation history	Window (last 5–10 messages)
Knowledge base	RAG context	Optional, for knowledge-based agents

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Build It, Use It, Ship It, Exercises, Key Terms, Common Pitfalls — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.