Agentic AI From Prediction to Action

Agentic AI: From Prediction to Action by Suresha H Parashivamurthy is an engineering guide (18 chapters, four parts) that traces the path from machine-learning fundamentals to building and governing autonomous AI agents in production. Its thesis is its title: the field has shifted from systems that predict - a model answers a question - to systems that act, pursuing goals, choosing their own steps, using tools, and changing the world. The first half is the foundation the second half recombines.

Part One covers the history of AI and the essentials of learning: linear algebra, loss functions, gradient descent, and neural networks with backpropagation. Part Two builds language understanding - tokenization, embeddings, the transformer and self-attention, and large language models with their scaling laws, context windows, and inference economics. Part Three bridges toward agents: generative families (VAEs, GANs, diffusion, multimodal), alignment and reasoning (RLHF, DPO, fine-tuning, inference-time compute), retrieval with vector and graph databases, and prompting, RAG, and tool use.

Part Four, the core, builds agents themselves: the perceive-think-act loop and ReAct pattern; planning, memory, tools, and context engineering; multi-agent topologies, frameworks, and the interoperability protocols MCP and A2A; data pipelines; and production concerns - reference architectures, model gateways, durability, security and sandboxing, LLMOps/AgentOps, observability, and token economics. It closes with evaluation, the "lethal trifecta" and prompt-injection defenses, governance, red-teaming, and future directions.

The book's recurring stance is sober rather than hyped. Its through-line: choose the least autonomy that solves the problem. Agents earn their complexity only when a task genuinely demands deciding its own steps; otherwise a fixed workflow is better engineering. The mature posture it advocates is a practical middle - powerful tools with real limits, bounded, observed, evaluated, and kept under human oversight. Its goal is not flawless autonomy, but trustworthy autonomy.