How do you prevent infinite agentic loops and runaway execution?

Short answer: Infinite agent loops are prevented by enforcing hard step counters, maximum token and execution time budgets, cyclic state detection (hashing past tool call history), human-in-the-loop approval thresholds, and deterministic state graph bounds (e.g. LangGraph recursion limits).

Detail

Autonomous agents operating in ReAct loops can get stuck repeatedly executing identical failing tool calls or oscillating between two unviable actions.

[Agent Execution Loop]
         │
         ├── Step Counter >= Max Steps (e.g. 10) ──► TERMINATE (Step Limit)
         │
         ├── Cumulative Token Spend >= Budget     ──► TERMINATE (Budget Exceeded)
         │
         └── Tool History Hash Duplicate Detected  ──► TERMINATE (Cyclic Loop)

Key Circuit Breakers

1. Max Recursion Depth: Set hard limits (e.g. recursion_limit = 15) in framework graph executors.
2. Action Hash Loop Detection: Maintain a hash set of (tool_name, tool_args) tuples executed in the current session. If the exact same call repeats 3 times consecutively with no state change, abort.
3. Financial / Token Caps: Track aggregate API token spend during the run; kill the process if spend exceeds $0.50.

Example

Python loop detector snippet:

import hashlib

class AgentCircuitBreaker:
    def __init__(self, max_steps=10, max_repeats=2):
        self.max_steps = max_steps
        self.max_repeats = max_repeats
        self.call_counts = {}
        self.current_step = 0

    def check_step(self, tool_name: str, tool_args: dict):
        self.current_step += 1
        if self.current_step > self.max_steps:
            raise RuntimeError("Circuit Breaker: Max execution steps reached.")

        # Create deterministic hash of tool call
        call_sig = f"{tool_name}:{str(sorted(tool_args.items()))}"
        call_hash = hashlib.md5(call_sig.encode()).hexdigest()

        self.call_counts[call_hash] = self.call_counts.get(call_hash, 0) + 1
        if self.call_counts[call_hash] > self.max_repeats:
            raise RuntimeError(f"Circuit Breaker: Cyclic loop detected on tool '{tool_name}'. Aborting.")

Interview tips

• Emphasize that production AI agents must fail gracefully, returning structured status diagnostic reports to the user.
• Discuss human-in-the-loop (HITL) checkpoints before executing destructive actions (e.g. database deletes or external emails).

⬅ Back to AI Agents and MCP · All topics