Q29: Autonomous Agent Guardrails: Runaway Loop Prevention, Quotas & Recursion Limits - AI Engineering & LLMs (50 Qs)

1. The Multi-Thousand Dollar Runaway Loop Incident

Without strict safeguards, autonomous agents can enter runaway loops:

An agent encounters an unexpected error and attempts 10,000 continuous retries overnight.
Two collaborating agents get stuck in recursive conversational ping-pong.
An agent loops continuously over a 5,000-file directory, burning hundreds of dollars in API tokens per hour.

2. Enterprise Safeguard Layers

vbnet

                              Agent Request
                                    │
    ┌───────────────────────────────┴───────────────────────────────┐
    ▼                               ▼                               ▼
[ Max Step Guard ]          [ Cost Budget Guard ]         [ Loop Cycle Detector ]
Max 15 iterations / task    Max $2.00 token spend / task  Identifies repetitive
Hard termination on breach  Hard termination on breach    action signatures

3. Production Multi-Layer Guardrail Implementation

python

class AgentBudgetGuard:
    def __init__(self, max_steps: int = 15, max_cost_usd: float = 2.00):
        self.max_steps = max_steps
        self.max_cost_usd = max_cost_usd
        self.current_steps = 0
        self.current_cost_usd = 0.0
        self.action_history = []

    def record_step(self, action_signature: str, token_cost_usd: float):
        self.current_steps += 1
        self.current_cost_usd += token_cost_usd
        self.action_history.append(action_signature)

        # 1. Step Quota Check
        if self.current_steps > self.max_steps:
            raise RuntimeError(f"Terminated: Exceeded maximum allowed step limit ({self.max_steps}).")

        # 2. Budget Quota Check
        if self.current_cost_usd > self.max_cost_usd:
            raise RuntimeError(f"Terminated: Exceeded financial cost budget (${self.max_cost_usd:.2f}).")

        # 3. Cycle Detection: Check if the last 3 actions are identical
        if len(self.action_history) >= 3 and len(set(self.action_history[-3:])) == 1:
            raise RuntimeError("Terminated: Detected infinite cyclic tool invocation loop.")