Anatomy of an AI Agent: How an AI agent works inside (The Professional and the Machine)

This is not a book about prompting. It is not a book about adoption strategy. It is the technical documentation of the internal architecture of a real agent application, read from the code. The ReAct loop, the provider router, memory, subagents, hooks, permissions, MCP, plugins — piece by piece, with diagrams and references to the exact file in the repository.The case study is SylvarSecDesktop, an Electron application owned by the author with ~55,000 lines of TypeScript, multi-provider LLM.What you'll find in this book29 technical chapters organized in 10 parts covering the complete anatomy of the agent runtime:The loop — ReAct from scratch: why a while and not a graph. Streaming deltas and events. The full API contract: messages, tools, finish_reason, thinking, ThinkingManager. Recovery from max_tokens, quotas, cancellation and resumption with snapshots.Providers — Provider router with the StreamCallbacks interface. Practical differences Anthropic vs OpenAI vs local LLMs. The "OpenAI-compatible" trap. Why a serious agent doesn't commit to one provider.Tools — The tool as a contract (schema vs handler). Registry and descriptions as invisible prompt. Streaming execution, parallelism and large result persistence. MCP: capabilities by protocol.Memory — The five layers: active messages, prompt cache, compaction, snapshots and persistent. AutoCompact, MicroCompact, agentMemory, autoDream. ContextAnalyzer as a measurement instrument. RAG as documentary memory.The team — Three child patterns: spawn, fork, InProcessTeammate. Swarm, A2A governance, task framework. Workflows with substitution and batch.Control, security and observability — Two-level permissions (tool and transport). Hooks, plan mode, guardrails and hook hardening. NDJSON to the frontend. VM sandboxing with VHDX and mount tables.Extensibility and UI — Slash commands, skills, versioned prompts. Plugin SDK and packs. Agent UI: rendering a live agent, voice, task notifications, XSS hardening.Cross-cutting vision and observability — Cost tracking, rate limiting, telemetry with session traces. Where agent architecture is heading.Appendices — Glossary, system prompt catalog, built-in tools catalog, master diagram of the runtime, references and how an agent is tested.Each chapter has six layersThe Question — What confusion this chapter resolves.The Diagram — The visible mechanics. ASCII and Mermaid, every arrow points to a repo file.How It Works — The mechanism step by step.The Decision — What SylvarSec chose, with alternatives and trade-offs.Example — Short and representative fragment.When It Does NOT Apply — Antipatterns and cases where the mechanism is surplus.Who this book is forAgent programmers who've spent weeks or months building and still don't feel like they understand what they're doing.AI architects who design multi-agent platforms and need to reference concrete mechanics (compaction, fork-vs-spawn, A2A) without ambiguity.Security engineers who audit agent applications and need to know where they break.Technical trainers who have to explain agents to others — developers, students, technical committees.CTOs and technical leaders who want to understand the mechanics before approving architecture.Book 12 of the series The Professional and the MachineCarlos Perez Gonzalez — AI Architect. Builder of production agent platforms.Juan C. Montes — Cybersecurity and infrastructure architect. Published in PHRACK #65. Read more