# Loop Engineering: Build an AI That Checks Itself in 25 Lines

> Source: <https://mer.vin/2026/11/loop-engineering-build-an-ai-that-checks-itself-in-25-lines/>
> Published: 2026-11-07 08:48:04+00:00

Give an AI one guess and it can be wrong. Put the same AI in a loop that checks each answer and sends back a hint, and it keeps going until it is right, or until the limit you set. This short guide builds that loop in **25 lines of free, open-source Python** with [PraisonAI](https://github.com/MervinPraison/PraisonAI).

```
%%{init: {"theme": "base", "themeVariables": {"background": "transparent", "lineColor": "#000000"}}}%%
graph TD
    T[Try: the AI guesses a number] --> C{Check: is it the secret?}
    C -->|yes| D[Stop: found it]
    C -->|no| F[Feedback: go higher or go lower]
    F --> L{Rounds left?}
    L -->|yes| T
    L -->|no| S[Stop: out of rounds]

    classDef agent fill:#8B0000,color:#fff
    classDef hook fill:#189AB4,color:#fff
    classDef decision fill:#444,color:#fff
    class T,D,S agent
    class F hook
    class C,L decision
```

Prefer to watch? The whole build is in a six-minute video: [Loop Engineering Made EASY: AI That Checks Itself](https://youtu.be/UO2teJGQZkk).

Please ⭐ star the repo[https://github.com/MervinPraison/PraisonAI](https://github.com/MervinPraison/PraisonAI)

## What is loop engineering?

You already own a loop: the thermostat. It checks the room, turns the heating on if it is too cold, checks again a few minutes later, and stops when it is warm enough. It never needs you to watch it.

An AI loop works the same way, with four parts: **try** (the AI answers), **check** (your code tests the answer), **feedback** (your code tells the AI what to fix) and **stop** (when it is right, or out of tries). Loop engineering is designing that loop, instead of typing a new message to the AI each time. The name comes from Addy Osmani’s post [Loop Engineering](https://addyosmani.com/blog/loop-engineering/) (7 June 2026).

## What you need

- Python 3.10 to 3.14 (free, from python.org). On Windows, tick “Add Python to PATH” in the installer.
- An OpenAI API key from platform.openai.com with a few dollars of credit. Each round is one paid request; a whole run cost me well under a cent.
- PraisonAI: `pip install praisonai` (on a Mac without a virtual environment, use`pip3` and`python3` ).

Tell your computer your key (this lasts for this terminal window only):

```
# Mac
export OPENAI_API_KEY="your-key-here"

# Windows PowerShell
$env:OPENAI_API_KEY="your-key-here"
```

## The code: guess_loop.py

Copy this into a plain-text file called `guess_loop.py`. Keep the spaces at the start of the lines.

``` python
from praisonai import Agent

SECRET = 38          # the number the AI must find
MAX_ROUNDS = 10      # the stop: never more than this

agent = Agent(
    name="Guesser",
    instructions="Guess my secret number from 1 to 100. Reply with one number only.",
    model="gpt-4o-mini",
    output="silent",
)
task = "Guess the number."

for round_no in range(1, MAX_ROUNDS + 1):
    reply = agent.start(task)                                  # 1. try
    guess = int("".join(c for c in reply if c.isdigit()) or 0)
    if guess == SECRET:                                        # 2. check
        print(f"Round {round_no}: {guess}   CHECK: correct!")
        print(f"DONE: found it in round {round_no}")
        break                                                  # 4. stop: it worked
    hint = "higher" if guess in range(SECRET) else "lower"
    print(f"Round {round_no}: {guess}   CHECK: wrong, go {hint}")
    task = f"{guess} is wrong. My number is {hint}. Guess again."   # 3. feedback
else:
    print(f"STOPPED: out of rounds (limit {MAX_ROUNDS})")      # 4. stop: out of rounds
```

| Lines | Part | What it does | 
|---|---|---|
| 3–4 | Setup | The secret number, and the limit: never more than 10 rounds | 
| 6–12 | Setup | The AI and its instructions: guess a number from 1 to 100, reply with one number only | 
| 15–16 | 1. Try | Ask the AI, then keep only the digits of its reply | 
| 17–20 | 2. Check + 4. Stop | If the guess is the secret, print it and stop | 
| 21–23 | 3. Feedback | Work out “higher” or “lower” and send it back as the next message | 
| 24–25 | 4. Stop | Out of rounds: stop, so it can never run forever | 

## Run it

```
python guess_loop.py
```

My run (5 November 2026, praisonai 4.7.23, gpt-4o-mini):

```
Round 1: 57   CHECK: wrong, go lower
Round 2: 34   CHECK: wrong, go higher
Round 3: 45   CHECK: wrong, go lower
Round 4: 40   CHECK: wrong, go lower
Round 5: 38   CHECK: correct!
DONE: found it in round 5
```

With one round (`MAX_ROUNDS = 1`) it was wrong every time. With 10 rounds it found 38 in 4 to 9 rounds over 8 runs. The AI guesses differently each time, but it remembers its earlier guesses, so each hint narrows it down.

## When the answer can’t be found

Change line 3 to `SECRET = 150` and line 4 to `MAX_ROUNDS = 5`. The AI was told 1 to 100, so it can never win:

```
Round 1: 50   CHECK: wrong, go higher
Round 2: 75   CHECK: wrong, go higher
Round 3: 87   CHECK: wrong, go higher
Round 4: 93   CHECK: wrong, go higher
Round 5: 96   CHECK: wrong, go higher
STOPPED: out of rounds (limit 5)
```

It stopped at round 5 in every run. That limit is the line that stops the bill.

*A loop without a stop rule is a bill without an end.*

`from praisonai import Agent` and `from praisonaiagents import Agent` both work. The Windows commands were not tested on Windows.

## Related on mer.vin

- [Harness Engineering: Build Your First AI Harness in 26 Lines](https://mer.vin/2026/11/harness-engineering-build-your-first-ai-harness-in-26-lines/)
- [Loop Engineering: Design Coding-Agent Systems Instead of Prompting Every Turn](https://mer.vin/2026/06/loop-engineering-design-coding-agent-systems-instead-of-prompting-every-turn/)
- [Karpathy Loop Engineering: Verifiers, State, and Bilevel Autoresearch](https://mer.vin/2026/07/karpathy-loop-engineering-verifiers-state-and-bilevel-autoresearch-explained/)