agentic-ai-kit 0.1.1

A progressive framework for building agentic AI systems with Google Gemini.

agentic-ai

A progressive Python framework for building agentic AI systems with Google Gemini.

agentic-ai turns common agentic AI patterns into reusable Python components: talking agents, memory, tool use, structured JSON output, reasoning workflows, supervisor delegation, parallel execution, debate, RAG, MCP tool servers, and applied data science agents.

This project started with a simple question: what happens if two agents talk and one of them is a dog that can bark? From that playful starting point, the patterns evolved into a reusable framework for building composable agentic systems.

---

Installation

pip install agentic-ai

For the latest GitHub version:

pip install --upgrade --force-reinstall git+https://github.com/shashank14581/agentic-ai-kit.git

Set your Gemini API key:

export GEMINI_API_KEY="your-key-here"

In Google Colab, store GEMINI_API_KEY in Colab Secrets and load it into os.environ.

---

Quick start

from agentic_ai.agents import BaseAgent

agent = BaseAgent(
    name="Alfred",
    sys_prompt="You are a witty British butler.",
    model="gemini-1.5-flash",
    extract_memory=False,
)

agent.think("Good morning! What should I do today?", stream=False)

---

What it supports

Capability	Component
Simple conversational agents	BaseAgent
Persona and short-term memory	BaseAgent, ShortTermMemory
LLM-extracted durable facts	BaseAgent.facts_store
SQLite-backed long-term memory	LongTermMemory
Shared state across agents	SharedMemory
Gemini function-calling tools	ToolAgent, ToolRegistry
Structured JSON output	JsonAgent
Reasoning / planning workflows	ReasoningAgent
Round-robin conversations	run_conversation
Supervisor delegation	run_supervisor
Parallel fan-out execution	run_parallel
Debate and judge pattern	run_debate
Chunking, embeddings, vector search	chunk_text, embed_texts, VectorStore
Retrieval-augmented generation	RAGAgent
MCP-style tool server and client	MCPServer, MCPClient
DataFrame analysis	AnalystAgent
Machine learning model training	MLEAgent
End-to-end auto modeling workflow	AutoModelAgent

---

Package layout

agentic_ai/
├── agents/
│   ├── base.py              # BaseAgent — memory, streaming, durable facts
│   ├── tool_agent.py        # ToolAgent — Gemini function-calling
│   ├── json_agent.py        # JsonAgent — structured JSON responses
│   ├── reasoning_agent.py   # ReasoningAgent — planning-style workflows
│   ├── analyst_agent.py     # AnalystAgent — dataframe profiling and analysis
│   ├── mle_agent.py         # MLEAgent — train and compare sklearn models
│   └── auto_model_agent.py  # AutoModelAgent — profile data, train model, summarize results
├── memory/
│   ├── short_term.py        # Sliding-window turn memory
│   ├── long_term.py         # SQLite-backed persistent memory
│   └── shared.py            # Thread-safe shared state across agents
├── tools/
│   ├── registry.py          # @tool decorator and ToolRegistry
│   └── builtins.py          # get_weather, add_numbers, search_wikipedia
├── patterns/
│   ├── orchestrator.py      # run_conversation, run_supervisor
│   ├── parallel.py          # run_parallel fan-out execution
│   └── debate.py            # run_debate with debaters and judge
├── rag/
│   ├── chunker.py           # chunk_text
│   ├── embedder.py          # embed_texts using Gemini embeddings
│   ├── vector_store.py      # In-memory cosine-similarity vector store
│   └── rag_agent.py         # RAGAgent ingest → retrieve → generate
└── mcp/
    ├── server.py            # MCPServer — HTTP tool server
    └── client.py            # MCPClient — calls a running MCPServer

---

Examples

All examples live in agentic_ai/examples/ and can be run directly:

File	Concept
01_talking_agents.py	Two agents in a round-robin conversation
02_personality_and_memory.py	Agent with persona and short-term memory
03_tool_agent_weather.py	Real weather tool through Gemini function-calling
04_json_agent.py	Structured JSON output
05_reasoning_agent.py	Planning-style reasoning workflow
06_supervisor.py	Supervisor delegates to specialist workers
07_parallel_agents.py	Fan-out: same task, multiple agents in parallel
08_debate.py	Two agents debate; a judge picks a winner
09_long_term_memory.py	SQLite-backed memory across sessions
10_rag_agent.py	Document ingestion and retrieval-augmented generation
11_mcp_demo.py	MCP tool server and agent integration

---

Data science agents

The kit also includes applied analytics and modeling agents.

from agentic_ai.agents import AutoModelAgent

agent = AutoModelAgent(
    project_context="Marketing audience selection and customer conversion modeling.",
    model="gemini-1.5-flash",
)

result = agent.run(
    df=df,
    target_col="converted",
    objective="Predict customer conversion.",
    interpret=False,
)

print(result["best_model"])
print(result["best_score"])

Use interpret=False when you want local sklearn training without an additional Gemini interpretation call.

---

Design philosophy

The framework separates reasoning from execution.

• Agents can reason, plan, route, critique, and summarize.
• Utilities can execute deterministic work such as memory lookup, chunking, vector search, dataframe profiling, and model training.
• Pattern functions provide lightweight orchestration.
• Specialist agents provide higher-level user-facing workflows.

This makes the library useful for both learning agentic AI patterns and building practical data science or automation workflows.

---

License

MIT