🐝 BeeAI Agent Systems — IBM Watsonx & Agentic AI with BeeAI Framework

---

📌 Project Overview

This project is a comprehensive, progressive exploration of Agentic AI using IBM's BeeAI Framework — one of the most advanced open-source frameworks for building production-grade AI agent systems.

Across 12 progressive tasks, the project builds from basic LLM calls to sophisticated multi-agent systems with specialized roles, tool orchestration, human-in-the-loop controls, custom tools, and RequirementAgent execution control — all powered by IBM Watsonx, Llama-4 Maverick, Granite, and GPT-5 Nano.

Domain: Agentic AI — BeeAI Framework
Language: Python 3.11
LLMs: IBM Granite 3.3 · Meta Llama-4 Maverick · OpenAI GPT-5 Nano
Platform: IBM Watsonx.ai

---

📂 Project Structure

BeeAI-Agent-Systems/
│
├── t1.py    # Watsonx.ai environment setup
├── t2.py    # Basic LLM chat — Granite 3.3 8B
├── t3.py    # Prompt templates with variables
├── t4.py    # Structured output with Pydantic
├── t5.py    # Minimal RequirementAgent (no tools)
├── t6.py    # Agent + WikipediaTool + Trajectory
├── t7.py    # Agent + ThinkTool + WikipediaTool
├── t8.py    # Controlled execution with Requirements
├── t9.py    # Force reasoning after every tool call
├── t10.py   # Human-in-the-loop (AskPermission)
├── t11.py   # Custom tool creation (Calculator)
├── t12.py   # Multi-agent travel planning system
└── README.md

---

🛠️ Tech Stack

Component	Technology
Agent Framework	BeeAI Framework
LLM 1	IBM Granite 3.3 8B Instruct (Watsonx)
LLM 2	Meta Llama-4 Maverick 17B 128E FP8 (Watsonx)
LLM 3	OpenAI GPT-5 Nano
Platform	IBM Watsonx.ai
Data Validation	Pydantic BaseModel
Memory	UnconstrainedMemory
Middleware	GlobalTrajectoryMiddleware
Tools	WikipediaTool · ThinkTool · OpenMeteoTool · HandoffTool
Async	asyncio

---

🚀 Progressive Learning Path — 12 Tasks

---

✅ t1.py — Environment Setup

Configure IBM Watsonx.ai credentials and project ID for Skills Network labs

---

✅ t2.py — Basic LLM Chat

llm = ChatModel.from_name("watsonx:ibm/granite-3-3-8b-instruct",
                           ChatModelParameters(temperature=0))
messages = [SystemMessage(...), UserMessage(...)]
response = await llm.create(messages=messages)

Simple chat with IBM Granite 3.3 8B — business idea brainstorming

---

✅ t3.py — Prompt Templates

# Custom mustache-style template engine
template = SimplePromptTemplate("Project: {{project_name}}, Problem: {{business_problem}}")
rendered = template.render({"project_name": "ML Classifier", ...})

Dynamic prompt rendering for data science project evaluation

---

✅ t4.py — Structured Output with Pydantic

class BusinessPlan(BaseModel):
    business_name: str
    elevator_pitch: str
    target_market: str
    revenue_streams: List[str]
    key_success_factors: List[str]

response = await llm.create_structure(schema=BusinessPlan, messages=messages)

GPT-5 Nano generates typed, validated business plans

---

✅ t5.py — Minimal RequirementAgent

agent = RequirementAgent(
    llm=llm,  # Llama-4 Maverick
    tools=[],  # No tools
    memory=UnconstrainedMemory(),
    instructions=SYSTEM_INSTRUCTIONS
)
result = await agent.run(ANALYSIS_QUERY)

Pure LLM cybersecurity analysis — no tools baseline

---

✅ t6.py — Agent + Wikipedia + Trajectory Tracking

agent = RequirementAgent(
    llm=llm,
    tools=[WikipediaTool()],
    middlewares=[GlobalTrajectoryMiddleware(included=[Tool])],
    requirements=[ConditionalRequirement(WikipediaTool, max_invocations=2)]
)

Research-enhanced agent with full tool usage tracking

---

✅ t7.py — ThinkTool + WikipediaTool

agent = RequirementAgent(
    tools=[ThinkTool(), WikipediaTool()],  # Reasoning + Research
    requirements=[
        ConditionalRequirement(ThinkTool, max_invocations=2),
        ConditionalRequirement(WikipediaTool, max_invocations=2)
    ]
)

Structured reasoning alongside Wikipedia research

---

✅ t8.py — Controlled Execution Requirements

requirements=[
    ConditionalRequirement(
        ThinkTool,
        force_at_step=1,      # Must think first
        min_invocations=1,    # At least once
        max_invocations=3,    # Maximum 3 times
        consecutive_allowed=False  # No repeated thinking
    ),
    ConditionalRequirement(WikipediaTool, ...)
]

Declarative control over tool execution order and behavior

---

✅ t9.py — Force Reasoning After Every Tool Call

ConditionalRequirement(
    ThinkTool,
    force_at_step=1,
    force_after=Tool,         # Think after EVERY tool call
    min_invocations=1,
    max_invocations=5,
    consecutive_allowed=False
)

Mandatory reasoning step after each tool invocation

---

✅ t10.py — Human-in-the-Loop (AskPermission)

requirements=[
    ConditionalRequirement(ThinkTool, force_at_step=1, ...),
    AskPermissionRequirement(WikipediaTool)  # Human approval required!
]

Production-ready security — human must approve external tool access

---

✅ t11.py — Custom Tool Creation

class SimpleCalculatorTool(Tool[CalculatorInput, ToolRunOptions, StringToolOutput]):
    name = "SimpleCalculator"
    description = "Performs basic arithmetic calculations"
    input_schema = CalculatorInput

    async def _run(self, input: CalculatorInput, ...) -> StringToolOutput:
        result = self._safe_calculate(input.expression)
        return StringToolOutput(f"Result: {result}")

Build custom BeeAI tools from scratch with Pydantic input validation

---

✅ t12.py — Multi-Agent Travel Planning System

# Agent 1 — Destination Research Expert
destination_expert = RequirementAgent(
    tools=[WikipediaTool(), ThinkTool()], ...)

# Agent 2 — Weather & Logistics Specialist
weather_agent = RequirementAgent(
    tools=[OpenMeteoTool(), ThinkTool()], ...)

# Agent 3 — Language & Culture Expert
language_expert = RequirementAgent(
    tools=[HandoffTool(...), ThinkTool()], ...)

# Coordinated multi-agent travel planning pipeline

3-agent specialized system with handoff coordination

---

🔑 BeeAI Concepts Mastered

Concept	Implementation
ChatModel	from_name() — Watsonx, OpenAI backends
RequirementAgent	Core BeeAI agent with tool + memory + requirements
ConditionalRequirement	Declarative tool execution control
AskPermissionRequirement	Human-in-the-loop approval
GlobalTrajectoryMiddleware	Full tool call tracking
UnconstrainedMemory	Persistent conversation memory
WikipediaTool	External research capability
ThinkTool	Structured reasoning steps
OpenMeteoTool	Real-time weather data
HandoffTool	Agent-to-agent communication
Custom Tool	Full BeeAI Tool class from scratch
Structured Output	create_structure() + Pydantic
Prompt Templates	Mustache-style variable rendering

---

🎓 Skills Demonstrated

• BeeAI Framework — complete agent architecture
• IBM Watsonx.ai integration — Granite + Llama-4 Maverick
• Progressive agent capability building (no tools → multi-agent)
• Declarative execution control with Requirements system
• Human-in-the-loop AI design with AskPermissionRequirement
• Custom tool development with Pydantic validation
• Multi-agent coordination with HandoffTool
• Tool usage trajectory tracking and middleware
• Structured LLM output with Pydantic BaseModel
• Custom prompt template engine
• Async Python — asyncio throughout
• Real-world use cases: Cybersecurity, Business Planning, Travel

---

📜 Certifications

Certification	Issuer	Platform
IBM Data Science Professional Certificate	IBM	Coursera
IBM Generative AI Professional Certificate	IBM	Coursera
IBM RAG and Agentic AI Professional Certificate	IBM	Coursera

---

🤝 Connect with Me