Deep Dive Series

Agentic AI Masterclass

Building Agentic Systems from Scratch in 2026

A build-it-yourself field manual for the engineers of autonomous systems

Curated by Arunabh Bhattacharjee · github.com/arubhatt

Updated 9 November 2026

Part I Foundations

01What an Agent Actually Is Agency, harnesses, and what changed by 2026.02The Model as a Component Thinking budgets, tool training, context limits.03Anatomy of an Agent Harness The reference architecture for the whole book.

Part II Orchestration & State

04The Agent Loop The canonical loop, from first principles.05Durable State & Resumption Surviving crashes without repeating side effects.06Planning, Decomposition & Task Graphs Turning a goal into executable structure.07Orchestration Patterns Composable shapes — and when not to use an agent.

Part III Context & Knowledge

08Context Engineering The window as a budget you spend deliberately.09Compaction & the Long Horizon Tasks that outlive a single context window.10Memory Architectures What the agent keeps, and what it must forget.11Retrieval & RAG for Agents Search as an action, not a preprocessing step.12Knowledge Graphs & Structured Memory When relationships are the answer.

Part IV Tools & Action

13Tool Calling from First Principles Schemas, errors, and the cost of a tool.14Tool Ecosystems at Scale MCP, discovery, and choosing among thousands.15Permissions, Sandboxing & Least Privilege Bounding what an agent can reach.16Code Execution as the Universal Tool Code as the highest-bandwidth action.

Part V Modality & the Human

17Executing on Multimodal Content Pixels, documents, audio, and artifacts.18Voice & Real-Time Agents Latency budgets, barge-in, and spoken tool calls.19Human Approvals, Escalation & Handoff Designing the seam between agent and person.

Part VI Safety & Failure

20Guardrails & Stop Conditions Bounding behavior, spend, and runtime.21Prompt Injection & Adversarial Agents The lethal trifecta and defense in depth.22Retries, Recovery & Failure Handling Failing safely in a non-deterministic system.

Part VII Measurement & Improvement

23Observability & Tracing Seeing inside a multi-step, non-deterministic run.24Evaluation Frameworks & Metrics Measuring outcomes and trajectories honestly.25Prompt Iteration & Feedback Loops Error analysis as an engineering discipline.26Self-Improvement Skills, tools, and memory the agent writes for itself.27Self-Learning Trajectories, distillation, and reinforcement learning.

Part VIII Multi-Agent & Production

28Multi-Agent Coordination Topologies, contracts, and when one agent wins.29From Prototype to Production Tenancy, cost, capacity, and progressive rollout.30Enterprise Governance & Compliance Identity, audit, residency, and the AI Act.

Part IX Appendices

AThe Reference Agent Capstone A 90-day plan to build the whole system.BReference Implementation Tour A guided read of the shipped code.