Limits, budgets and stopping

Three ways to stop, and you want all of them

A step limit alone is not enough. An agent can burn a large budget inside a small number of steps, or loop between two tools that keep undoing each other while making perfect progress by its own account.

Guard	Stops
step limit	an unbounded loop
cost budget	expensive steps inside the step limit
no-progress detector	a loop that repeats the same call forever

The third is the one people leave out. An agent calling search("refund policy") for the fourth time is not working, however confident its reasoning sounds, and the cheapest detector is simply noticing the repeat.

Your turn: build Guard with max_steps, max_cents and a repeat limit. check(tool_name, args, cents) records a step and returns None to continue or a string naming which limit was hit.

class Guard:
    def __init__(self, max_steps=6, max_cents=50, max_repeats=2):
        self.max_steps = max_steps
        self.max_cents = max_cents
        self.max_repeats = max_repeats
        self.steps = 0
        self.cents = 0
        self.seen = {}

    def check(self, tool_name, args, cents):
        """Record a step. Return None to continue, or the reason to stop."""
        return None