The Agent Loop in Action

Observe, decide, act, check

The loop is simple. (1) Observe: run the tests and see what fails. (2) Decide: read the relevant code and form a hypothesis. (3) Act: make a small edit. (4) Check: run the tests again. (5) Stop when the checks pass, or when a limit is reached. The example below is a miniature, real agent loop: a throwaway project has a bug in apply_discount (it divides the percentage by 10 instead of 100); the harness provides run_tests, read_file and edit_file tools; and a scripted stand-in plays the model, proposing the five actions a real model would plausibly choose. Nothing here calls a real model, so it shows the mechanics (tools, observations, checks, a hard step cap), not the intelligence. A real agent chooses its own actions, which is exactly why the verification step matters.

A miniature coding agent, run

I ran this with plain Python 3 (standard library only), using a throwaway project created in a temporary folder. Step 1 runs the tests and finds test_discount_ten_percent failing. Steps 2 and 3 read the file and apply an exact-match edit changing / 10 to / 100. Step 4 re-runs the tests, which now pass, and step 5 finishes with a summary. A final independent test run confirms the fix. The model here is a fixed script, so this demonstrates the harness, not a real model.

import os, tempfile, textwrap

def make_project(root):
    files = {
        "shop/__init__.py": "",
        "shop/pricing.py": textwrap.dedent("""
            def apply_discount(price, percent):
                \"\"\"Return price after a percentage discount.\"\"\"
                return price - price * percent / 10

            def add_tax(price, rate=0.18):
                return round(price * (1 + rate), 2)
        """),
        "shop/cart.py": textwrap.dedent("""
            from shop.pricing import apply_discount, add_tax

            class Cart:
                def __init__(self):
                    self.items = []

                def add(self, name, price):
                    self.items.append((name, price))

                def total(self, discount_percent=0):
                    subtotal = sum(p for _, p in self.items)
                    return add_tax(apply_discount(subtotal, discount_percent))
        """),
        "tests/__init__.py": "",
        "tests/test_pricing.py": textwrap.dedent("""
            import unittest
            from shop.pricing import apply_discount, add_tax

            class PricingTests(unittest.TestCase):
                def test_discount_ten_percent(self):
                    self.assertEqual(apply_discount(200, 10), 180)

                def test_discount_zero(self):
                    self.assertEqual(apply_discount(200, 0), 200)

                def test_tax(self):
                    self.assertEqual(add_tax(100), 118.0)
        """),
    }
    for rel, text in files.items():
        path = os.path.join(root, rel)
        os.makedirs(os.path.dirname(path), exist_ok=True)
        with open(path, "w") as f:
            f.write(text.lstrip("\n"))

import subprocess, sys, re

def run_tests(root):
    p = subprocess.run([sys.executable, "-m", "unittest", "discover", "-s", "tests", "-t", "."],
                       cwd=root, capture_output=True, text=True)
    failed = re.findall(r"^(?:FAIL|ERROR): (\S+)", p.stderr, re.M)
    return {"passed": p.returncode == 0, "failed": failed}

def read_file(root, rel): return open(os.path.join(root, rel)).read()

def edit_file(root, rel, old, new):
    path = os.path.join(root, rel); text = open(path).read()
    if text.count(old) != 1: return "error: old text must appear exactly once"
    open(path, "w").write(text.replace(old, new)); return "ok"

# A scripted stand-in for the model: it proposes the next action given what it has observed so far.
SCRIPT = [
    ("run_tests", {}),
    ("read_file", {"rel": "shop/pricing.py"}),
    ("edit_file", {"rel": "shop/pricing.py", "old": "price * percent / 10", "new": "price * percent / 100"}),
    ("run_tests", {}),
    ("finish", {"message": "Fixed apply_discount: percent was divided by 10 instead of 100."}),
]

def agent(root, max_steps=8):
    script = iter(SCRIPT)
    for step in range(1, max_steps + 1):                       # a hard cap on steps
        name, args = next(script)
        if name == "run_tests": obs = run_tests(root)
        elif name == "read_file": obs = "(%d chars)" % len(read_file(root, **args))
        elif name == "edit_file": obs = edit_file(root, **args)
        else:
            print(f"step {step}: finish -> {args['message']}"); return
        print(f"step {step}: {name}({', '.join(f'{k}=...' for k in args)}) -> {obs}")

with tempfile.TemporaryDirectory() as root:
    make_project(root)
    agent(root)
    print("final test run:", run_tests(root))

Output:

step 1: run_tests() -> {'passed': False, 'failed': ['test_discount_ten_percent']}
step 2: read_file(rel=...) -> (200 chars)
step 3: edit_file(rel=..., old=..., new=...) -> ok
step 4: run_tests() -> {'passed': True, 'failed': []}
step 5: finish -> Fixed apply_discount: percent was divided by 10 instead of 100.
final test run: {'passed': True, 'failed': []}

Start every run from green

Make sure tests pass before the agent starts; otherwise you cannot tell which failures it caused.

Quick check: Why does the loop end with an independent test run?

• Because the model cannot read files
• To make the run longer
• Tests are optional decoration
• The agent's own claim of success is not proof; tests are