Your first agent

Let us build a real agent from scratch. It will guess a person's age from clues. By the end you will have written a file and run it against a real AI model.

Step 1: Make a file

Create a file called age.nt in any folder. Put this inside it:

agent age
  description: Guesses a person's age from clues.
  model: anthropic/claude-sonnet-5
  instructions: |
    Guess the person's most likely age from the clues you are given.
    Reply with your best single guess and one short reason.
  input:
    clues: string
  output:
    age: number
    reason: string

Here is what each line means, in plain words:

• agent age gives your agent a name: age.
• description is a short note about what it does.
• model is which AI brain it uses.
• instructions is what you want it to do. Write it like you are talking to a helpful assistant.
• input is what you will give it (here, some clues, which are text).
• output is what you want back (a number age and a text reason).

Step 2: Check the file

Before running, ask NT to check your file for mistakes:

nt validate

If everything is fine, you will see a success message. If not, NT tells you what to fix.

Step 3: Run it

Set your AI key, then run the agent with some clues:

export ANTHROPIC_API_KEY=sk-ant-...

nt run age -i '{"clues":"bought the first iPhone at 22"}'

The -i part is the input you are giving it, written as {"clues": "..."}.

You will get an answer back like this:

{ "age": 39, "reason": "The first iPhone came out in 2007." }

That is a working agent.

What just happened

You described an agent in plain text, NT connected it to a real AI model, and it answered your question in the exact shape you asked for.

From here you can give your agent more powers.

Next