Chapter 2 of 13
The loop, by hand
The loop is five verbs: observe, decide, act, append, repeat. There is no framework in that sentence and there does not need to be, because the whole thing is about twenty lines of Python. The part that surprises people is where the state lives. It is the transcript, a plain list that grows by one entry per step, and it is the only memory the agent has. That list is also re-sent to the model on every turn, so a 40-step run pays for roughly 800 entries of context rather than 40, which is why a step cap is a budget control and not only a safety rail. Loops die three ways in production. Nothing reads the stop signal, so the agent keeps working after it is done. The model names a tool that does not exist and a naive dispatcher raises, killing the run mid-task. Or the model never signals done at all. Each has a specific fix, and this lab runs all three failures and proves the guarded loop survives every one.
The lab: read it, then run it
#!/usr/bin/env python3
"""
LAB AE2: The loop, by hand, and the three ways it dies.
The loop is observe, decide, act, append, repeat. The transcript IS the state: a
list that grows by one entry per step, and the only memory the agent has. Three
things kill a loop in production: nothing reads the stop signal, the model names
a tool that does not exist, and the model never signals done at all. Each has a
specific fix, and this lab runs all three and proves the guarded loop survives.
Run: python3 modules/academy-content/labs/agentic-engineering/ae2-loop-by-hand.py
"""
import sys, os
_cands = [os.path.join(os.path.dirname(__file__), "..") if "__file__" in globals() else None,
os.path.join(os.getcwd(), "..", "labs"), os.path.join(os.getcwd(), "labs")]
for _c in _cands:
if _c and os.path.exists(os.path.join(_c, "academy_llm.py")):
sys.path.insert(0, os.path.abspath(_c)); break
from academy_llm import tool_route
REGISTRY = {
"weather": lambda city: {"paris": 22, "tokyo": 30}[city.strip().lower()],
"calculator": lambda expr: sum(int(x) for x in expr.replace("+", " ").split()),
}
def act(tool, arg):
"""FIX for death 2: an unknown tool or a thrown error becomes an OBSERVATION
the loop can read, never an exception that escapes and kills the run."""
if tool not in REGISTRY:
return "ERROR: no tool %r. available: %s" % (tool, sorted(REGISTRY))
try:
return REGISTRY[tool](arg)
except Exception as e:
return "ERROR: %s failed: %s" % (tool, e)
def run(goal, script, max_steps, honor_stop=True):
"""The whole loop. `script` stands in for the model's chosen calls; a real
agent parses these out of the model's output each turn."""
transcript = ["GOAL: " + goal] # state = a list that only grows
answer, steps, ended_by = None, 0, "step cap"
while steps < max_steps: # FIX for death 3: the cap
tool, arg = script[steps] if steps < len(script) else ("weather", "paris")
steps += 1
if tool == "final": # the model's stop signal
if honor_stop: # FIX for death 1: somebody reads it
answer, ended_by = arg, "stop signal"
transcript.append("step %d: STOP with %r" % (steps, arg))
break
transcript.append("step %d: STOP ignored" % steps)
continue
obs = act(tool, arg)
transcript.append("step %d: %s(%r) -> %r" % (steps, tool, arg, obs))
return answer, steps, ended_by, transcript
GOAL = "the temperature in Paris plus 10"
GOOD = [("weather", "paris"), ("calculator", "22 + 10"), ("final", "32")]
print("DEATH 1: nothing reads the stop signal")
a, s, why, _ = run(GOAL, GOOD, max_steps=6, honor_stop=False)
print(" unguarded: %d steps, answer %r, ended by %s" % (s, a, why))
a1, s1, why1, t1 = run(GOAL, GOOD, max_steps=6, honor_stop=True)
print(" guarded : %d steps, answer %r, ended by %s" % (s1, a1, why1))
d1 = (s == 6 and a is None) and (s1 == 3 and a1 == "32" and why1 == "stop signal")
print("")
print("DEATH 2: the model calls a tool that does not exist")
try:
REGISTRY["database"]("paris") # what a naive dispatcher does
except KeyError as e:
print(" naive dispatch raised KeyError(%s) and the run would be over" % e)
a2, s2, why2, t2 = run(GOAL, [("database", "paris")] + GOOD, 6)
print(" guarded : %d steps, answer %r, ended by %s" % (s2, a2, why2))
print(" recovery line -> %s" % t2[1])
d2 = a2 == "32" and any("ERROR" in line for line in t2)
print("")
print("DEATH 3: the model never signals done")
a3, s3, why3, t3 = run(GOAL, [("weather", "paris"), ("calculator", "22 + 10")], 4)
print(" guarded : %d steps, answer %r, ended by %s" % (s3, a3, why3))
d3 = a3 is None and s3 == 4 and why3 == "step cap"
print("")
print("transcript grew to %d entries, one per step plus the goal" % len(t2))
print("routing still works on the ask: %r" % tool_route("what is the weather in Paris", list(REGISTRY)))
ok = d1 and d2 and d3
print("")
print("LOOP TERMINATED CORRECTLY IN ALL THREE DEATH MODES: %s" % ("YES" if ok else "NO"))
if not ok:
sys.exit(1)
print("Stop rule, error-as-observation, step cap. Next: the tools it calls.")
Runnable lab
ae2-loop-by-hand.pyImplement the loop in about twenty lines, then run it into all three death modes and prove each one terminates correctly.
Proves: LOOP TERMINATED CORRECTLY IN ALL THREE DEATH MODES: YES
Runs in your browser via Pyodide. First run loads the runtime once; no install, no server.
Read the three results against each other. With nothing reading the stop signal the loop burned all 6 steps and returned no answer, while the same loop honoring the signal finished in 3 steps with 32. The hallucinated database tool became an ERROR observation at step 1 instead of a KeyError, and the loop read that line, moved on, and still answered 32 by step 4. The script that never signals done hit the cap at 4 steps and reported that it ended on the cap with no answer, which is the correct outcome and not a bug: an agent that stops and says it failed is worth far more than one that invents a result to have something to return. Every one of those fixes is a few lines. The loop is easy. The tools it calls are where the engineering is, and that is next.
Check your understanding
1. What is an agent's state during a loop?
2. What should a dispatcher do when the model calls a tool that does not exist?
3. Why is a step cap a cost control and not only a safety rail?