Jean Tessier’s Reviews > AI Agents in Action, Second Edition: Intelligent workflows with LLMs, MCP, A2A, and more > Status Update

Deployment options: run in the browser, run in a service, run in a queue worker. Use Docker. Manage permissions. Build a threat model based on what you're doing and what date it uses.

Test-driven agent development: code agents by manipulating LLM instructions and prompts. Use agents to check the work of other agents: evaluators, grounding, critics. Critics use rubrics, elaborate prompts that tell the agent what to look for and how to rate it. Phoenix, Langfuse, or LangSmith to capture telemetry (and OpenTelemetry) and diagnose performance and errors.

The bit trying to relate agent "memory" to human cognitive neuroscience didn't land for me. Not sure why agent would care about "sensory memory" as opposed to image/video just being a different format of data. A lot of the integration is done in prompts, but the discussion lacks the subtleties of telling the LLM when to use what tool. (I guess that changes quickly as models evolve.)

RAG, search, memory: vector db are great for semantic search (similarity). Use an embedding model to converts pieces of data into vectors. Vector search usually looks at angle, not scale, only if they point in the same direction. RAG can combine from multiple sources and rerank when using hybrid search. Easy setup: ChromaDB (vector DB), server-memory MCP (graph DB). Context = short-term, external = long-term.

Instructing the agent how to think: chain-of-thought CoT for single pass and ReAct for iterative. Tree-of-through does breadth-first search, Reflexion does depth-first. They consume lots of tokens. Sequential Thinking MCP is a scratch pad for the agent, not thinking itself. Good examples with tools and MCP. Reflexion needs to know the answer, so it's weird. Maybe for drafting plans??

Splitting one agent doing three things into three agents, each doing one thing. Guardrails as deterministic functions or agents themselves. Agents are pieces of code inside the app, akin to services in Grails. Wiring can be in prompts or explicit in the control code.

Patterns for coordinated agents: flow, the hub-and-spokes, hierarchy. Patterns for communication flow. Decision making, control, and communication. Too high level, I want to see actual setup and decision logic and dispatch mechanisms.

Separate tools in a separate module and access then using MCP. Keeps told isolated and reusable. Keeps agent simple.

Different ways to run MCP servers. List of available servers for talking to Google or Slack.

MCP fundamentals. STDIO vs SSE, er, I mean Streamable HTTP (book came out last month and is already out-of-date).