Tool Abuse

An autonomous agent with access to tools is dangerous if not constrained.

The Infinite Loop Problem

If an agent cannot solve a problem, it often tries the same action repeatedly.

# A naive agent loop
while True:
    cmd = agent.think()
    if cmd == "list_files":
        output = os.listdir()
        # Agent sees file, forgets, lists again...

Destructive Actions

Without a graph structure defining permitted transitions, an agent might decide the best way to "clean up logs" is rm -rf /logs.

Scenario: The Runaway Agent

1. Agent tries to debug code.
2. Agent decides to "edit config".
3. Agent breaks config.
4. Agent loses network access.
5. Agent loops 10,000 times trying to "ping google" to restore access.

LangGraph prevents this by enforcing Recursion Limits and Checkpoints where a human can step in.