智能体与智能体架构导论

Authors: Alan Blount, Antonio Gulli, Shubham Saboo,Michael Zimmermann, and Vladimir Vuskovic Content contributors Enrique ChanMike ClarkDerek Egan Curators and editorsAnant Nawalgaria Designer Table of contents Table of contents The Orchestration Layer22Core Design Choices23Instruct with Domain Knowledge and Persona23Augment with Context24Multi-Agent Systems and Design Patterns24Agent Deployment and Services26Agent Ops: A Structured Approach to the Unpredictable27Measure What Matters: Instrumenting Success Like an A/B Experiment29Quality Instead of Pass/Fail: Using a LM Judge29 Table of contents Agents are the natural evolutionof Language Models, made useful From Predictive AI toAutonomous Agents Artificial intelligence is changing. For years, the focus has been on models that excel atpassive, discrete tasks: answering a question, translating text, or generating an image froma prompt. This paradigm, while powerful, requires constant human direction for every step. This new frontier is built around AI agents. An agent is not simply an AI model in a staticworkflow; it's a complete application, making plans and taking actions to achieve goals. Itcombines a Language Model's (LM) ability toreasonwith the practical ability toact, allowing it to handle complex, multi-step tasks that a model alone cannot. The critical capability is thatagents can work on their own, figuring out the next steps needed to reach a goal without a This document is the first in a five-part series, acting as a formal guide for the developers,architects, and product leaders transitioning from proofs-of-concept to robust,production-grade agentic systems. While building a simple prototype is straightforward, •Core Anatomy:Deconstructing an agent into its three essential components: the •A Taxonomy of Capabilities:Classifying agents from simple, connected problem-solvers •Architectural Design:Diving into the practical design considerations for each •Building for Production:Establishing the Agent Ops discipline needed to evaluate,debug, secure, and scale agentic systems from a single instance to a fleet with Building on the previousAgents whitepaper1andAgent Companion2; this guide providesthe foundational concepts and strategic frameworks you will need to successfully build, Words are insufficient to describe how humans interact with AI. We tend toanthropomorphize and use human terms like “think” and “reason” and “know.” We don'tyet have words for "know with semantic meaning" vs "know with high probability of In the simplest terms, an AI Agent can be defined as the combination of models, tools, anorchestration layer, and runtime services which uses the LM in a loop to accomplish a goal. •The Model (The "Brain"):The core language model (LM) or foundation model that servesas the agent's central reasoning engine to process information, evaluate options, andmake decisions. The type of model (general-purpose, fine-tuned, or multimodal) dictates •Tools (The "Hands"):These mechanisms connect the agent's reasoning to the outsideworld, enabling actions beyond text generation. They include API extensions, codefunctions, and data stores (like databases or vector stores) for accessing real-time, factual •The Orchestration Layer (The "Nervous System"):The governing process thatmanages the agent's operational loop. It handles planning, memory (state), and reasoning Chain-of-Thought4orReAct5) to break down complex goals into steps and decide whento think versus use a tool. This layer is also responsible for giving agents the memory •Deployment (The "Body and Legs"):While building an agent on a laptop is effective forprototyping, production deployment is what makes it a reliable and accessible service.This involves hosting the agent on a secure, scalable server and integrating it with At the end of the day, building a generative AI agent is a new way to develop solutions tosolve tasks. The traditional developer acts as a "bricklayer," precisely defining every logicalstep. The agent developer, in contrast, is more like a director. Instead of writing explicit code You'll quickly find that an LM's greatest strength—its incredible flexibility—is also your biggestheadache. A large language model's capacity to doanythingmakes it difficult to compel it todoone specific thingreliably and perfectly. What we used to call “prompt engineering” andnow call “context engineering” guides LMs to generate the desired output. For any single Debugging becomes essential when issues arise. "Agent Ops" essentially redefines thefamiliar cycle of measurement, analysis, and system optimization. Through traces and logs,you can monitor the agent's "thought process" to identify deviations from the intended critical components: domain expertise, a defined personality, and seamless integrationwith the tools necessary for practical task completion. It's crucial to remember that When an agent is precisely configured with clear instructions, reliable tools, and a