生成式AI 401：智能体（底层技术）

potential really came to liferecentlywith the release of products like Claude Code/Cowork, OpenAl Codex/Agent Builder,and OpenClaw.In thisprimer, given the amount oftechnicalities,wetrytopeel offthejargon and packaging ofagentic systems andexplainwhat is going on under the hood.Al agents are built on top of the basic LLM models and are likelyto become an important mark.moerdler@bernsteinsg.comFiroz Vallji, CFA +19173448316firoz.vallji@bernsteinsg.comShelly Tang, CFA partofthe application layer of the Al software tech stack.Compared toa basic LLMthatcannotremember,has stale information,cannotperformactions andhallucinatesknowledge, context, data, tools, rules and guidelines.The easiest way to understand how agents work is this: During inference,the basic LLM shelly.tang@bernsteinsg.com does not have any knowledge beyond its training data and user input, so how to supplytheLLMmorecontexttohelp itsucceed at atask?The answeristo dumpeverything itneeds to know in the input including: user information, business context, how to break downthe steps, what tools are required and how to use them, and relevant data such as login,password, customer and order details etc. This is known as the context. information in the input soon feels overwhelming as the variety of tasks and complexityscale.Tobetter organize andoptimize the context supplied during each task,engineershaveparsed outparts ofthecontext into differentmodules.Namely,short-termmemory,long-termmemory,Tools and rules/Workflows/Skills. These agentic components are evolving fast and might morph into something elsetomorrow. However, the key themes underpinning this era of Gen Al developmentis that 1)we aremoving from building inside an LLMto building around an LLM,which means thatagentic Al could unlock model capability but does not imply model improvement; 2) we aremovingfroma moregeneralizedapproachtoamorespecializedapproachto enhancetask performance. The amount of design and engineering choices that go into building anagent, webelieve,signals that step-function breakthroughsmight take longerthan duringthe pretraining scaling era. We are excited about the future of Agentic Al, and believe that by finding the perfectbalancebetween determinism (consistency)andnon-determinism (flexibility)foreachtypeofworkflow,agentic Al can open upa lotmore newuse casesfor software.Atthe sametime, it is important to recognize the amount of complexity and work that goes into settinglong contextperformance andmulti-agent communications,webelievethedevelopmenttimeline for customized agents will take time, longer than many think and will require moreexperts (thus the need forForward Deployed Engineers) Agentic Al represents the next wave of innovation in Generative Al technology and is likely to become an important part of theapplication layer of the Al software tech stack (versus LLM's andagentic development tools whichwill become part of thePaaSlayer). We will see more agents to be rolled out at the application software companies in our coverage. We believe havingagentic Al could require a lot more efforts than many imagined so it is likely to take longer. Unlike pure LLM where compute is predominantly GPU-based, many of the agentic steps will be performed on the CPU's. Asagentic Al adoption increases we should see an increase in CPU consumption-a possible tailwind forthe hyperscalers suchas Microsoft and Oracle within our coverage, driving both additional revenue and incrementally higher gross margins for Alworkloads overtraining or pure inferencing. Building effective agents also require significant groundwork in data infrastructure. Hence, we see many names in ourwill also use more data driving an incremental tailwind for Cloud database vendors such as Microsoft, Oracle, MongoDB andSnowflake. discussed in Forward Deployed Engineering (FDE): Where Al software meets the real world. How did wego from the2022 ChatGPT to agents today?.A very stripped-down way of looking at agents..7Memory: Keeping track of history9Memory: Retrieve additional knowledge..11Tools and MCP: Enable actions....12Agent Tool schemaWhat is an MCP?..15How LLMs know when and which tool to call...17Workflow/skills/systemprompts:Improveaccuracyandefficiency19 capabilitywouldbe in the Generative Al evolution.Althoughthe concept ofagents oragentic Al startedto popularizearoundlate2023/2024, its significanceand potential reallycametolifemorerecentlywiththe release of products likeClaudeCodeClaude Cowork,OpenAlCodex,OpenAlAgentBuilder,and OpenClaw. Whilepeopleare amazedat thenew capabilities of these agentic systems and the speedat whichthey innovate,most ofthenewproducts andtheassociated terminologies arenotthemostuser-friendlytothe non-developers/technicalcohorts.Wetherefore trytopeeloff the jargon andpackaging ofthe agentic Al system inthis primer, exposethebackbones and explainwhat is going on under the hood. Our motivation to do so is two-fold. First, as agentic Al is likely going to become a major