👷 What is an agent
An agent is a Jarvis worker: a component that receives a task, carries it out, and returns a result. Think of a real company—you wouldn’t hire one person to “do everything.” You hire someone to support customers, someone to sell, someone to organize documents. Jarvis works the same way: each agent is like an employee with a clear role. The service says what needs to be done; the agent is who does.
🧩 Agent = employee, not the entire system
The system (Jarvis) is the company. The service is the department. The agent is the person who sits in the chair and does the work. When you separate these three things, it’s easy to design the solution: first the department (service), then who works in it (agents), then what each person needs to do their job well.
✓ An agent is
- ✓A worker with a role
- ✓Who executes the service’s “how”
- ✓Something that receives input and returns output
- ✓One component of the system, not the whole system
✗ An agent isn't
- ✗All of Jarvis in a single box
- ✗A jack-of-all-trades with no focus
- ✗Just "a clever prompt"
- ✗Something that decides the strategy on its own
🎯 Specialized: one task each
The golden rule for agents: one agent, one job. An agent that tries to do everything does everything badly—it loses focus, the instructions become a jumble, and no one can tell whether it got things right. A specialized agent, on the other hand, is easy to instruct, test, and improve. Instead of one “super agent,” Jarvis uses several small agents, each honed for its task.
✗ A jack of all trades
- ✗Long, confusing instructions
- ✗Gets a little wrong about everything
- ✗Hard to know where it failed
- ✗Impossible to measure success
✓ Multiple specialized agents
- ✓Short, clear instruction for each agent
- ✓Each one gets its part right
- ✓Easy to see who failed
- ✓You can measure and improve them one by one
💡 Practical tip
If you can’t describe an agent’s job in one sentence ("it classifies emails as urgent or normal"), it’s doing too much. Split it into two. The shorter the description, the better the agent.
📝 Instructions and Context
A new worker needs two things on day one: the instruction (what to do and how) and the context (information about the company, the client, the situation). It's the same with an agent. The instruction describes the work; the context provides the lay of the land. An instruction without context leads to a generic response; context without an instruction becomes information without direction. Together, they make the agent work like someone on the inside.
The instruction
"You answer customer questions about delivery. Be direct and kind, and never promise a delivery time that isn't in the table."
The context
The table of deadlines, exchange policies, store name, and brand tone—the landscape the agent navigates to make decisions.
The task
The concrete question that comes up in the moment: "when will my order 1042 arrive?" Instruction + context + task = the right answer.
💡 Practical tip
When an agent answers “incorrectly,” it’s almost never the instruction—it’s the missing context. Before rewriting the instruction, ask: did it have the information it needed to get it right?
🧠 Agent memory and rules
In addition to instructions and context, a good agent has memory (what it remembers from the conversation or history) and rules and limits (what it can and must never do). Memory prevents the client from repeating everything in every message. Rules protect the company: an agent without limits is a risk. This is where you write the agent's "profile" — and it fits in just a few key:value lines.
An agent’s profile (key:value)
Illustrative profile — every part of the agent fits on one line. Memory and context are covered in more depth in Module 2.5.
🧠 Memory
What the agent carries from one message to the next: the customer’s name, the order in question, and what has already been said. Without this, it "forgets" and the customer has to repeat everything.
⛔ The rules and limits
The lines an agent never crosses: don't make up data, don't promise what you can't deliver, and don't act beyond your authority. "No" protects as much as "yes."
✅ Success criteria
How do you know if an agent is doing a good job? You need an success criterion: a clear statement of what “success” means for that agent. Without a standard, any response seems acceptable, and you never know whether it’s worth improving. With a standard, you can test, compare, and improve the agent—and confidently say whether it’s ready.
✗ Vague criterion
- ✗"The agent responds well"
- ✗"The customer is satisfied"
- ✗Can’t be measured
- ✗Everyone evaluates differently
✓ Clear criteria
- ✓"Gives the correct order status"
- ✓"Without opening a ticket for a human"
- ✓You can verify them case by case
- ✓Everyone agrees on what “success” means
🎯 The criterion comes before measurement
On Day 3, we'll measure real results. But measurement only works if the criterion was written down beforehand. Defining “what success means for this agent” is part of the design—not something you discover afterward.
🧭 Which intent the agent fulfills
Here, the module closes with the question that comes before everything else: what intent does this agent fulfill? Don’t start by creating the agent; start with the intention it serves. The system’s intention (its soul, from 2.1) unfolds into services, and each service into agents. Each agent is the fine point of a larger intention. If you can’t say which intention an agent fulfills, that agent probably shouldn’t exist.
💡 The question to ask before choosing a tool
The temptation is to start by asking, “What agent should I create?” The intent architect starts with, “What intent do I need to fulfill?”—and only then does the agent emerge as a natural consequence. Intent is the cause; the agent is the effect.
The intent
"I don't want any customer to go without an answer about their delivery." That's the pain point that justifies the system.
The service
Support for deliveries—the department that handles this intent.
The agent
The agente_de_entregas—the fine point that puts that intention into practice, with its complete profile.
🧾 A checklist for a healthy agent
- Has a clear intent that justifies its existence.
- Has only one job — can be described in one sentence.
- Has instructions, context, memory, rules, and limits — the complete profile.
- Has a success criterion verifiable, written in advance.
Self-check (optional): what question comes BEFORE creating an agent?
🎯 Module summary
Next module:
2.4 — Skills: what each agent can do in practice.