Book Summary
Book 2 of The Agentic Frontier. The Amazon listing names ReAct state machines, multi-agent topologies, a tri-tier memory engine with GraphRAG, MCP tool servers, and deterministic schema checks. Kindle and Kindle Unlimited.

The book

The Agentic Frontier: Volume 2 is Book 2 of 3 in The Agentic Frontier by Vatsal Shah. The Kindle edition was published on 4 September 2026. Language: English. ISBN-13 978-9334485738. ASIN B0HHSD1NP8.

The full listing title is: The Agentic Frontier: Volume 2: Agentic Architecture: ReAct State Machines, Multi-Agent Swarms, GraphRAG, MCP & Sovereign Fleets.

Open the Kindle edition on Amazon.

What is inside

The description on the Amazon page names these blocks:

  1. Act-observe-reflect engine — ReAct loops, state-machine transitions, and token-budget pruning.
  2. Multi-agent topologies — hierarchical supervisors, peer-to-peer swarms, and deadlock recovery in LangGraph, CrewAI, and AutoGen.
  3. Tri-tier memory and GraphRAG — working memory, vector stores, and Neo4j entity graphs.
  4. Model Context Protocol — tool servers that connect agents to databases, APIs, and machines over JSON-RPC.
  5. Schema checks and sandboxes — deterministic schema enforcement and zero-trust sandboxes.

Who it helps

Read it for the role you are in. Each block on the listing maps to a job.

Engineer. Start with the ReAct loop and the state machine, then the memory tiers.

Architect. Use the topology section to choose a supervisor or a swarm, and the MCP section to keep tools vendor-agnostic.

Security reviewer. Use the sandbox and schema section before an agent can call a tool.

Engineering lead. Use Volume 1 for the operating picture, and this volume for the build.

How to use it

Work the blocks in the order the listing names them.

  1. Write the loop. Act, observe, reflect, with a token budget.
  2. Pick a topology. Supervisor or swarm, and how a deadlock ends.
  3. Split the memory. Working scratchpad, vector store, entity graph.
  4. Attach tools through MCP. One server, JSON-RPC, no vendor lock in the tool layer.
  5. Enforce the schema. Reject a bad payload before the sandbox runs it.

What you leave with

  • A loop shape you can implement.
  • A topology choice with a recovery path.
  • A memory split, not one bucket.
  • A tool boundary through MCP.
  • A schema check in front of the sandbox.

The series on this site

Questions about using this in a team workflow: contact me.

Vatsal Shah

Vatsal Shah

AI Leader · Solution Architect · TPM

I design autonomous AI systems, enterprise architectures, and publish deep technical content for global organisations.