This repository contains the presentation slides and architectural documentation for the technical workshop titled "Autonomous AI Agents and Tool Integration: From Text Completion to Autonomous Action", presented at the Şarkışla Science, Art, and Culture Club.
It covers the architectural evolution of Large Language Models (LLMs) from passive text generators to goal-oriented autonomous systems capable of reasoning, planning, and interacting with external APIs, databases, and environments.
- Paradigm Shift (Chatbot vs. AI Agent):
- Contrasting passive, pre-trained parametric models with goal-driven agents that interact with external environments.
- Anatomy of an AI Agent:
- Brain (LLM Engine): Core reasoning, semantic parsing, and orchestration.
- Memory Architecture: Short-term memory (Context Window) vs. Long-term memory (Vector Databases / RAG).
- Planning: Task decomposition, iterative sub-goal generation.
- Tools & Actuators: REST APIs, SQL engines, bash terminals, and environment runtimes.
- The Golden Rule: "Models Do Not Execute Code; They Generate Parameters":
- The LLM acts as an orchestrator emitting structured JSON payloads; code execution and security enforcement remain strictly on the backend.
- Function Calling & Tool Contracts:
- Defining rigid tool interfaces using JSON Schema / OpenAPI specifications and managing the end-to-end execution loop.
- ReAct Framework (Reason + Act):
- The Thought $\rightarrow$ Act $\rightarrow$ Observation $\rightarrow$ Final Answer cycle for grounded, evidence-based problem solving.
- Error Recovery & Self-Correction:
- Treating runtime system/API exceptions as new observations to dynamically pivot execution paths instead of failing.
- Guardrails & Safety Controls:
- Bounding agent recursion with step limits (Max Iterations).
- Token buffers and financial thresholds (Budget Guardrails).
- Human-in-the-Loop (HITL) approval gates for state-mutating, irreversible operations (POST/DELETE/Execute).
Presentation.pdf: The minimalist, high-contrast, annotated presentation slide deck used during the session.LICENSE: MIT License.
- Speaker: Himmet Can Umutlu
- Host Institution: Şarkışla Science, Art, and Culture Club
- Duration: ~25–30 Minutes
All documentation and presentation assets in this repository are distributed under the MIT License. You are free to fork, adapt, and use this material for non-commercial educational workshops or community technical sessions.