How Do I Explain "Agent" to My Wife?
From 'the Lark bot can't do math' to 'raising your own AI lobster' — an AI-agent explainer for normal humans. Thirteen burning questions covering LLMs, tokens, Tools, MCP, RAG, Skills, Memory, Multi-Agent systems, and 2026 model prices. The AI isn't dumb — it just hasn't been raised properly yet.
中文版 / Chinese Version: 怎么向你老婆解释什么是 Agent？
From “the Lark bot can’t do math” to “raising your own AI lobster” — an agent explainer for normal humans
💬 How It Started
Last night, my wife suddenly said:
“My boss asked me to look into Lark’s AI bot — whether we can add it to our group chats to track client requirements, contracts, payments, collections, sync everything to Lark Docs, and run settlement reports.”
Sounds like a reasonable ask. She tried it, then came back:
“I added the AI to the group, asked it to add up some numbers — and it got them wrong! Isn’t AI supposed to be amazing? It can’t even handle arithmetic?”
What followed was a barrage of soul-searching questions—
🤔 Q1: Isn’t AI Supposed to Be Powerful? Why Does It Fail at Basic Math?
Her confusion: Adding up 3 payments is elementary-school math. How does AI get it wrong?
Answer: Great question — it goes straight to the nature of today’s AI.
ChatGPT, Ernie, the AI inside Lark — at their core they’re all a thing called a Large Language Model (LLM).
Here’s the key: an LLM is essentially a master of word completion, not a calculator.
It works like this:
You say: 1+1=
It thinks: based on the trillions of words I've read,
the most likely thing after "1+1=" is "2"
It answers: 2Notice — it didn’t calculate the answer. It guessed it.
It’s like a liberal-arts student who has read every math textbook: most simple problems it can “recite” correctly, but ask it to compute the weighted average collection period of 47 receivables and it starts making things up. Because it has never truly computed anything.
🎯 One-liner: an LLM is a language genius and a math illiterate. It works by guessing the next word, not by computing.
🤔 Q2: Wait — How Does It “Guess”? What’s a Token?
Her follow-up: You say it guesses the next word — how, exactly? And you programmers keep saying “token.” What is that?
Answer: Let’s take these apart.
How does a large model work?
Imagine a genius friend who has read every word on the internet — news, papers, novels, chat logs, code, encyclopedias… trillions of words.
She didn’t memorize every sentence. She developed intuition — give her the start of any passage, and she can continue it convincingly.
That’s the training process:
Training data (trillions of words) → learn language patterns → an "intuition" modelWhen you ask it something:
Your question → the model emits one word at a time by intuition → the answerSo what’s a token?
A token is the model’s smallest unit of text processing.
Roughly: one token ≈ a word, or part of a word.
| What you typed | Tokens |
|---|---|
| ”今天天气真好" | "今天” / “天气” / “真” / “好" |
| "Hello World" | "Hello” / ” World" |
| "Payment 50000 CNY" | "Payment” / ” 50000” / ” CNY” |
The model guesses one token, appends it, guesses the next — until it decides to stop.
Why should you care about tokens?
Because AI costs money, billed by the token.
Like phone calls billed by the minute:
- What you send the AI → input tokens (costs money)
- What the AI replies → output tokens (also costs money, usually more)
Prices vary wildly between models — the cheapest domestic Chinese models process a million tokens for pennies; expensive imported ones cost hundreds of RMB. Details later.
🎯 One-liner: the token is the AI’s smallest unit of text — and its billing unit. Models “guess” answers one token at a time.
🤔 Q3: If LLMs Can’t Do Math, Why Not Give Them a Calculator?
Wife (very reasonably): If it gets math wrong, just hand it a calculator. Problem solved?
Answer: Congratulations — that thought is precisely the most important breakthrough direction in AI over the past two years.
Exactly right. The solution is: give the AI Tools.
AI before tools:
User: total up these 5 payments for me
AI: (guessing wildly) Hmm… about 48,350?
User: That's wrong!AI with tools:
User: total up these 5 payments for me
AI: (thinks) This needs precise arithmetic. Calling the calculator.
AI → calculator tool: 12000 + 8500 + 15000 + 6800 + 9200
Calculator → returns: 51500
AI: The total of these 5 payments is 51,500.That is the concept of Tools.
Whatever the AI is bad at, it delegates to an external tool:
| AI is bad at | The tool for it |
|---|---|
| Math | Calculator |
| Fresh information | Search engine |
| Reading files | File reader |
| Sending messages | Email / messaging API |
| Querying databases | SQL tool |
| Drawing | Image tool |
Think of the AI as a brilliant brain with no hands — it knows what should be done but needs “hands” to do it. Tools are its hands.
🎯 One-liner: Tools turn AI from “all talk” into “getting things done.” Calculator for math, search for facts — the AI’s job is choosing the right tool.
🤔 Q4: How Do Tools Get Plugged In? What’s MCP?
Wife: You say you “install” tools on the AI — how, concretely? You people keep mentioning “MCP Server.” What is that?
Answer: This question hits the bullseye.
There are infinitely many tools in the world — calculators, databases, Lark, WeChat, email, calendars — each with a different way of being used. If every AI had to integrate with every tool separately, chaos.
So the industry built a unified standard: MCP (Model Context Protocol).
An analogy:
Before MCP:
Like every country having a different power outlet —
you travel with a bag of adapters.
AI wants Lark? Write a Lark-specific integration.
AI wants WeChat? Write another one.
AI wants your database? Yet another…
Exhausting.With MCP:
Like the whole world standardizing on USB-C.
Charging, data, display — one port does it all.
Any tool that ships an MCP Server can be plugged straight into any AI.Concretely:
- MCP Server = the tool’s “USB-C port,” exposing its capabilities in a standard format
- MCP Client = the AI side’s “USB-C port,” which knows how to call those capabilities
- MCP Protocol = the agreed rules of communication
MCP was open-sourced by Anthropic (the Claude company) in late 2024. The standard proved so useful that OpenAI, Google, Microsoft, and Amazon all announced support. By 2026, MCP SDK downloads passed 97 million per month, with over 13,000 public servers on GitHub — the de facto standard for AI tool connectivity, USB-C style. In late 2025, Anthropic donated MCP to the Linux Foundation’s Agentic AI Foundation, co-governed by Anthropic, OpenAI, and Block — fully public infrastructure now.
So when my wife wants the AI to read Lark docs and write Lark sheets, what she needs is a Lark MCP Server — exposing Lark’s capabilities (read docs, write sheets, send messages) in the standard format. Lark has officially open-sourced lark-cli: 200+ commands covering 11 business domains — docs, sheets, calendar, mail — built precisely for AI agents.
🎯 One-liner: MCP is the universal port standard connecting AI to tools. Adopted by OpenAI, Google, Microsoft, and Amazon, with ~100M monthly downloads — it’s infrastructure now.
🤔 Q5: So LLM + Tools = “Agent”?
Wife: So large model + tools = agent?
Answer: Close — but one crucial thing is missing: autonomous decision-making.
Compare:
Ordinary AI chat (ChatBot)
You ask → it answers → done
You ask again → it answers again → doneQuestion-answer, one at a time. The AI never initiates anything.
An Agent
You say: organize this month's receivables for all clients,
build the report, send it to Finance
The Agent works autonomously:
1. 🤔 Think: I need this month's chat records first
2. 🔧 Tool: read the Lark group history
3. 🤔 Think: now extract the receivables information
4. 🔧 Tool: parse the chats into structured data
5. 🤔 Think: data extracted; compute the totals
6. 🔧 Tool: calculator for the statistics
7. 🤔 Think: now generate the report
8. 🔧 Tool: create a Lark sheet, fill in the data
9. 🤔 Think: finally, send it to Finance
10. 🔧 Tool: send via Lark to the Finance colleague
Agent: Done — the report went to Xiao Wang in Finance. Want to take a look?See the difference?
Agent = LLM (brain) + Tools (hands) + autonomous planning and execution (initiative)
It doesn’t wait for step-by-step commands. You give it a goal; it decomposes the task, picks tools, executes step by step, and adjusts when it hits problems.
As a formula:
🤖 ChatBot = LLM
🦾 Agent = LLM + Tools + planning + a loopThat’s what the word Agentic means — giving AI initiative: from passively answering to actively doing.
🎯 One-liner: an agent isn’t “AI that chats” — it’s AI that thinks, acts, and adjusts on its own. Brain + hands + initiative = Agentic.
🤔 Q6: Why Does the AI Forget What I Told It?
Wife (her most exasperated question): I told it our company’s settlement rules! Ask it again later and it has no idea! How is this not stupid?
Answer: The classic. Everyone who uses AI hits this.
The reason is simple: LLMs are born without memory.
In every conversation, what the AI can “see” is a limited window called the context window.
┌──────────────────────────────────────────────┐
│ Context window (say, 1M tokens) │
│ │
│ System prompt │
│ What you said earlier (if it still fits) │
│ The AI's earlier replies (if they still fit) │
│ Your newest question │
│ │
│ ⚠️ Anything beyond the window → gone. │
└──────────────────────────────────────────────┘Good news: the window grew enormously in two years — from 128K tokens in 2024 to 1M tokens in 2026 for GPT-5.5, Claude Opus 4.7, and DeepSeek V4 (roughly 7–8 novels), with Google Gemini 3.1 Pro at 2M.
Bad news: however big, it’s finite — and every new conversation starts empty. Everything from the last chat: wiped.
Like a person with only short-term memory — they follow everything while you’re talking, but earlier content gets squeezed out as you go. Worse: tomorrow they remember nothing of today.
The fix?
Agents solve it by attaching an external memory system.
AI without memory:
every conversation → starts from zero → remembers nothing
Agent with memory:
learns something important → writes it to a memory file
next conversation starts → reads the memory file → "Ah, right!"Like humans and notebooks — the brain forgets, the notebook doesn’t.
| Memory type | Analogy | Example |
|---|---|---|
| Short-term | The current conversation | This chat’s context |
| Long-term | A notebook | Important info saved to files/DB |
| Working | Sticky notes | Temporary notes for the current task |
So my wife’s problem, at root: that Lark AI bot has no long-term memory configured — it understood the settlement rules in the moment and forgot them by the next conversation.
🎯 One-liner: LLMs are born amnesiac — every conversation starts from zero. Agents fix it with external memory (files, databases).
🤔 Q7: Can the AI Just Go Read Our Company Docs?
Wife (lightbulb moment): Instead of me feeding it rules one by one, can it go read our documents itself? The settlement rules and client info are all in Lark Docs.
Answer: That idea of yours is one of the hottest techniques in AI right now — RAG (Retrieval Augmented Generation).
Intimidating name, dead-simple principle:
AI without RAG
You: what's our settlement cycle with Client A?
AI: Sorry, I don't have information about your company's business.The AI only knows public knowledge from training. Your internal docs? Never seen them.
AI with RAG
You: what's our settlement cycle with Client A?
The AI's internal flow:
1. 🔍 Search the knowledge base for "Client A settlement cycle"
2. 📄 Find the relevant passage in "Client Settlement Rules v3.2"
3. 📖 Stuff that passage into the context window
4. 🤔 Answer based on the document
AI: Per "Client Settlement Rules v3.2", Client A settles monthly net-30 —
bills submitted by the 5th, payment completed by the 15th.RAG in essence: search first, answer second.
Analogy:
- AI without RAG = an employee who works purely from memory of the company handbook
- AI with RAG = an employee who checks the filing cabinet before answering
How it’s implemented:
Step 1: feed company documents into the system
Lark docs, Excel sheets, PDF contracts…
→ chopped into passages
→ converted into a fast-searchable format (vectorized)
→ stored in a "knowledge base"
Step 2: when the user asks
Question → search the knowledge base for the most relevant passages
→ send passages + question to the AI together
→ the AI answers grounded in real materialThat’s why virtually every AI platform — Coze, Dify, FastGPT — makes the knowledge base a core feature. Without RAG, the AI is a generic assistant flying blind; with RAG, it becomes a specialist that knows your business.
🎯 One-liner: RAG lets AI “check the files before answering” instead of guessing from memory. Feed your docs into a knowledge base and it becomes your business expert.
🤔 Q8: What Are Skills, Then? How Are They Different From Tools?
Wife: You also keep saying Skills. Isn’t that the same as Tools?
Answer: Different — but easily confused. Analogy:
Tools are individual capabilities:
- A hammer drives nails
- A wrench turns bolts
- A calculator computes
Skills are a complete methodology for a class of tasks:
- The “renovation skill” = knowing when to hammer, when to wrench, what order to work in
- The “bookkeeping skill” = knowing how to gather data, which formulas, what report format
In AI terms:
| Tools | Skills | |
|---|---|---|
| What | One concrete capability | A complete workflow |
| Analogy | A hammer | Knowing how to renovate |
| Example | Read file, send message, compute | ”Extract client needs from chats and build the report” |
| Contains | One API call | Prompts + multiple tools + execution steps |
A Skill is a higher-level concept — it tells the AI “for this kind of task, follow this playbook,” and the playbook uses various Tools.
For instance, a “client reconciliation” Skill:
Skill: client reconciliation
Steps:
1. Read payment records from the Lark group chat
2. Extract amount, date, client name
3. Compare against the receivables ledger in Lark Sheets
4. Flag discrepancies, compute outstanding amounts
5. Generate the reconciliation report, send to the right colleagues
Tools used: Lark message reader, sheet read/write, calculator, message sender🎯 One-liner: a Tool is one move; a Skill is a whole playbook. Skill = prompt + tool combo + workflow.
🤔 Q9: Put It All Together and That’s a Complete Agent?
Wife: I think I’m getting it… summarize for me?
Answer: Let’s string the whole line together:
A complete Agent =
🧠 Brain (LLM)
↓ understands, reasons, decides
🔧 Tools
↓ executes concrete actions (compute, search, read/write files…)
🔌 Tool interface (MCP)
↓ plug-and-play for any tool
📚 Knowledge base (RAG)
↓ consults company docs and business material
📋 Skills
↓ predefined workflows and methodology
🧠 Memory
↓ remembers what matters — no more amnesia
🔄 Autonomous loop (Agent Loop)
↓ think → act → observe → think again… until doneBack to my wife’s need — client information management plus settlement statistics. What she needs is exactly this agent:
- Brain: a solid model (GPT-5.5, Claude Opus 4.7, DeepSeek V4)
- Tools: Lark doc read/write, sheet ops, messaging, calculator
- Interface: MCP to Lark (lark-cli is already open source)
- Knowledge base: feed in the settlement rules, client info, contract templates
- Skills: how to extract client needs and contract info from chats, how to run settlement stats
- Memory: past reconciliation results, client preferences, frequent contacts
The built-in Lark AI she was using probably has only the “brain” and none of the rest — hence wrong math (no calculator tool), forgotten rules (no memory), and business ignorance (no knowledge base).
🎯 One-liner: complete agent = brain + tools + interface + knowledge base + skills + memory + loop. Missing any one piece breaks it.
🤔 Q10: What If One Agent Can’t Keep Up? Can You Run Several?
Wife: This agent thing sounds capable — but if the work is heavy and messy, can one agent really handle it?
Answer: Great question. Just as one person can’t do a whole company’s work and you need a team — multiple agents can collaborate too.
This is one of the hottest trends of the past two years: Multi-Agent systems.
Suppose my wife runs a complex monthly settlement:
Single-agent mode
One agent reads chats AND does the math AND builds reports AND sends notices…
Like one person being accountant, clerk, statistician, and messenger at once —
exhausting and error-prone.Multi-agent collaboration
📋 Orchestrator Agent (project manager): decomposes tasks, assigns specialists
↓
🔍 Data-collection Agent: extracts all payments and contracts from Lark chats
↓
🧮 Computation Agent: verifies amounts, computes totals, flags discrepancies
↓
📊 Report Agent: produces the properly formatted settlement report
↓
📨 Notification Agent: sends the report, pings pending items
📋 Orchestrator: everyone's done — here's the consolidated result.Each agent focuses on what it does best — a well-drilled small team.
To let agents talk to each other smoothly, Google introduced the A2A (Agent-to-Agent) protocol in 2025 — now also donated to the Linux Foundation with 150+ supporting organizations. If MCP is “the USB-C port connecting AI to tools,” A2A is “the walkie-talkie channel between AIs.”
| Protocol | Solves | Analogy |
|---|---|---|
| MCP | How agents use tools | USB-C port (device connection) |
| A2A | How agents collaborate | Walkie-talkie channel (person to person) |
🎯 One-liner: Multi-agent systems bring teamwork to AI — division of labor, each to their strength. MCP connects tools; A2A connects agents.
🤔 Q11: What’s This “Raising a Lobster” Everyone Talks About?
Wife: People online keep saying they’re “raising an AI” or “raising a lobster.” What does that mean? How?
Answer: Ha — “raising a lobster” just means training and cultivating your own AI agent.
Why “raising”? Because it really is like onboarding a new employee:
Fresh out of the box
You: write my weekly report
AI: Certainly! Here is your weekly report: this week's work included…
(a wall of template filler)
You: that's not the style at all!After some raising
You: weekly report
AI: On it — your preferred style:
3 key wins this week first, then next week's plan,
in your team's usual Lark Docs format,
casual tone but exact numbers.
Done — want any changes?“Raising” is really:
- Give it an identity — tell it who it is (e.g., you’re my work assistant; style: concise and direct)
- Give it memory — let it retain your preferences, habits, go-to information
- Give it knowledge — feed company docs and business rules into RAG
- Give it skills — teach it your recurring tasks (Skills)
- Give it tools — connect the external capabilities it needs (Tools + MCP)
- Keep giving feedback — praise what’s good, correct what’s not
Exactly like onboarding a junior hire — you don’t expect a newcomer to know all the house rules on day one, but you teach, and they keep getting better.
🎯 One-liner: “raising a lobster” = continuously training your AI — identity, memory, knowledge, skills, tools, feedback — until it truly gets you.
🤔 Q12: And How Much Does All This Cost?
Wife (pragmatist edition): I get the theory. What’s the bill?
Answer: The good news of 2026 — AI is 10 to 100x cheaper than two years ago, and still falling.
The core cost is model token consumption. Prices vary hugely; here’s the April 2026 snapshot:
International models (per million tokens, at 1 USD ≈ 7.25 CNY)
| Model | Input | Output | Notes |
|---|---|---|---|
| GPT-5.5 | $5 (¥36) | $30 (¥218) | OpenAI’s latest flagship, 1M context, native agent capabilities |
| GPT-4.1 nano | $0.10 (¥0.73) | $0.40 (¥2.9) | OpenAI’s cheapest; fine for simple tasks |
| Claude Opus 4.7 | $5 (¥36) | $25 (¥181) | Anthropic’s latest, 1M context, strongest at coding |
| Claude Sonnet 4.6 | $3 (¥21.8) | $15 (¥109) | The value pick; enough for most scenarios |
| Gemini 3.1 Pro | $2 (¥14.5) | $12 (¥87) | Google’s flagship, 2M context |
Chinese models (per million tokens)
| Model | Input | Output | Notes |
|---|---|---|---|
| DeepSeek V4-Flash | ¥1 | ¥2 | Open source, 1M context, the value king for daily tasks |
| DeepSeek V4-Pro | ¥12 | ¥24 | Open source, 1.6T parameters, approaching Claude Opus level |
| Qwen3.5-flash | $0.10 (¥0.7) | $0.40 (¥2.9) | Alibaba, 1M context, absurdly cheap |
| Qwen3.5-plus | $0.40 (¥2.9) | $2.40 (¥17.4) | Flagship capability, still very low price |
| Qwen3.6-Plus | Free in preview | Free in preview | Newest flagship, SWE-bench 78.8%, 1M context |
| Doubao Seed 2.0 Pro | $0.47 (¥3.4) | $2.37 (¥17.2) | ByteDance, multimodal, 200M users |
What do these numbers mean?
Concretely: my wife has the AI process 50 group messages a day, extract client info, and run one settlement pass.
- Roughly 50K tokens/day (input + output)
- On DeepSeek V4-Flash: about ¥0.005/day (half a fen)
- On Qwen3.5-flash: about ¥0.009/day (under one fen)
- On Claude Sonnet 4.6: about ¥0.3/day
- Per month: a few mao to a few RMB on domestic models; tens of RMB on imported flagships
Conclusion: for everyday office work, AI usage cost is now effectively negligible.
🎯 One-liner: 2026 AI is absurdly cheap. A month of everyday office use on a domestic model costs less than a bottle of water.
🤔 Q13: So How Do I Actually Get One of These?
Wife: Fine, I’m convinced. How do I do it?
Answer: Depends on the level of solution you want. From simple to complex:
Option 1: Use a Ready-Made Enterprise AI Assistant (Zero Barrier)
Use the AI built into Lark, DingTalk, or WeCom.
- Lark AI: built-in doc summarization, sheet analysis, meeting minutes; extensible via the Coze platform. Lark open-sourced lark-cli — 200+ commands across 11 business domains
- DingTalk AI: launched Agent OS with the general-purpose “Wukong” agent, executing thousands of enterprise functions, full MCP support
- WeCom AI: Tencent open-sourced wecom-cli — AI agents can drive messages, calendars, docs, meetings, and more across 7 categories
Pros: works out of the box. Cons: limited flexibility for complex scenarios. Cost: usually bundled with the enterprise subscription.
Option 2: Build on an AI Platform (No Code)
Assemble your own agent on Coze, Dify, or FastGPT.
Coze — from ByteDance, now at version 2.5: pick your model, wire tools, design workflows; integrates with Lark, WeChat, DingTalk; knowledge base (RAG), memory, and skills built in; drag-and-drop, no code.
Dify — open-source AI app platform: visual workflow builder + agent framework + RAG; supports virtually every mainstream model; self-hostable so your data stays yours; free and open source, good for teams with some technical footing.
FastGPT — knowledge-base Q&A specialist: excels at turning company documents into an AI-queryable knowledge base; handles Word, PDF, Excel, web pages; perfect for the “make AI understand our business” scenario.
Pros: flexible, covers most scenarios. Cons: takes time to configure and tune. Cost: platform often free; models billed by token (cheap, as covered).
Option 3: Build It Yourself (Maximum Flexibility)
Call model APIs directly; write your own tools and workflows.
Mainstream frameworks:
- OpenAI Agents SDK — official, MCP support, GPT-5.5 has native agent workflows
- Google ADK — from Google, A2A protocol support
- LangChain / LangGraph — the most popular open-source frameworks, richest ecosystem
- Claude Code — Anthropic’s terminal coding agent; completes development tasks autonomously from the command line
Pros: total customization. Cons: requires engineers. Cost: mostly model API tokens.
My Recommendation
For my wife’s scenario, Option 2 fits best — build an agent on Coze or Dify, connect Lark, load the knowledge base, teach it the settlement rules. No code, a usable version in a day or two, then keep “raising” it.
🎯 One-liner: solutions run from zero-barrier to fully custom. Most people are fine with Coze or Dify — no code, up in a day or two.
📝 The Very End
Back to the opening scene — my wife wants AI to manage client information and run settlement statistics. Entirely achievable.
The key insight: today’s AI is not an omnipotent god — it’s a “digital employee” you configure and cultivate.
- Gets math wrong → install a calculator (Tools)
- Doesn’t know your business → let it read the docs (RAG)
- Forgets things → give it a notebook (Memory)
- Can’t reach Lark → give it the port (MCP)
- Won’t work proactively → make it an agent (Agent)
- One can’t keep up → give it a team (Multi-Agent)
- Too expensive → switch to a domestic model (DeepSeek V4, Qwen — pennies a day)
That is what Agentic means — moving AI from passive answering to active doing.
From 2025’s “Year of the AI Agent” into 2026, the field has outpaced everyone’s expectations. MCP passed 97M monthly downloads and became infrastructure; DeepSeek V4 approaches the strongest closed models at 1/7th the price; GPT-5.5 and Claude Opus 4.7 shipped in the same week fighting for the crown; Alibaba’s Qwen3.6-Plus presses the front rank on coding benchmarks; domestic model prices raced down to one RMB per million tokens…
All of it says one thing: AI agents are no longer programmers’ toys — they’re productivity tools for everyone.
Next time my wife asks “why is this AI so dumb,” I can say:
“It’s not dumb — it just hasn’t been raised yet. Like a new intern: teach it, give it tools, let it take notes, give it files to consult, and it gets good. And raising one costs a few mao a month — cheaper than a bottle of water.”
If this article helped you understand agents, share it with an equally confused friend.