AI Agents
LLMs, Agentic Workflows, and Tool Use
Preface
This collection of lecture notes and practical guides focuses on AI agents, with an emphasis on Large Language Models (LLMs), LLM reasoning, agentic workflows, and tool integration, including the development of reasoning-centric LLMs from scratch.
The materials consolidate insights from both research and practice, covering the design, orchestration, and evaluation of LLM-powered agents across diverse domains.
Each section begins with learning objectives, core concepts, and references, followed by illustrative examples and hands-on frameworks.
The notes emphasize AI design patterns, LLM reasoning strategies, Retrieval-Augmented Generation (RAG), and multi-agent systems, offering both conceptual depth and implementation guidance. They are not intended to replace formal textbooks, but to complement them through practical, framework-driven insights for building scalable, trustworthy, and adaptive LLM agents.
The intended audience is expected to have a foundational understanding of programming and familiarity with the following LLM-related concepts and frameworks:
-
AI / LLM design patterns, including:
- LLM reasoning and Chain-of-Thought (CoT) prompting
- Tree-of-Thought (ToT), Graph-of-Thought (GoT), and reflection-based reasoning
- Retrieval-Augmented Generation (RAG)
- Agent orchestration workflows and planner–executor design
- Memory and context management (short-term, episodic, and long-term)
- Self-reflective, hierarchical, and multi-agent collaboration patterns
- LLM reasoning and Chain-of-Thought (CoT) prompting
-
Frameworks and libraries:
- LangChain – composable pipelines for tool-augmented reasoning
- LangGraph – graph-based agent orchestration and reasoning flow control
- CrewAI – collaborative and role-based multi-agent coordination
- MCP (Model Context Protocol) – cross-model interoperability and context exchange
- LangChain – composable pipelines for tool-augmented reasoning
-
Evaluation and benchmarking
- Methods, metrics, and standardized benchmarks for LLMs and AI agents
- Benchmarks: AgentBench, ToolBench, ChatEval, SWE-Bench
- Evaluation dimensions: reasoning correctness, factuality, coherence, tool-use accuracy, and ethical alignment
- Methods, metrics, and standardized benchmarks for LLMs and AI agents
-
Multimodal and tool-integrated AI
- Integrating vision, text, and speech into agentic workflows
- Connecting APIs, databases, and external environments for tool use
- Designing interactive and embodied agents for perception and control
- Integrating vision, text, and speech into agentic workflows
-
Safety, ethics, and alignment
- Robustness, trustworthiness, and ethical safeguards for agentic systems
- Defenses against adversarial inputs, data poisoning, and prompt injection
- Responsible deployment and human-in-the-loop design
- Robustness, trustworthiness, and ethical safeguards for agentic systems
-
Emerging frontiers
- LLM + DRL: reinforcement learning agents with reasoning and planning
- LLM + GNNs: structured and relational reasoning for decision systems
- LLM compilers and world models: translating reasoning into executable plans
- AI operating systems: multi-agent ecosystems and autonomous orchestration
- LLM + DRL: reinforcement learning agents with reasoning and planning
Citation
Kumar, K. (2025). AI Agents: LLMs, Agentic Workflows, and Tool Use. Edition 2025-11.
License
This work is licensed under the MIT License.