How to detect tool call loops in AI agents

[bmdhodl]

A common failure mode in autonomous agents: the LLM decides to call the same tool with the same arguments over and over. It gets a result, doesn't know what to do with it, and tries again. And again. And again.

This is different from a retry — the agent genuinely thinks each call is a new idea. The model has no memory that it already tried this.

What a loop looks like

Your agent is supposed to research a topic. Instead:

1. Calls web_search("AI safety research papers") → gets results
2. Calls web_search("AI safety research papers") → same results
3. Calls web_search("AI safety research papers") → same results
4. ... burns through your API budget

Catching it at runtime

AgentGuard's LoopGuard watches for this pattern and kills the agent before it compounds:

from agentguard import LoopGuard, LoopDetected

# Trigger after 3 identical calls in a window of 6
guard = LoopGuard(max_repeats=3, window=6)

# In your agent loop, check each tool call:
for tool_name, tool_args in agent_tool_calls:
    try:
        guard.check(tool_name, tool_args)
    except LoopDetected as e:
        print(f"Loop detected: {e}")
        break
    # ... execute the tool call

The sliding window means the agent can call the same tool multiple times with different arguments (that's productive work). It only triggers when the exact same (tool, args) pair repeats.

Fuzzy loops

Sometimes the agent varies the arguments slightly — search("weather NYC") then search("NYC weather"). For this, FuzzyLoopGuard tracks tool-level repetition regardless of arguments:

from agentguard import FuzzyLoopGuard

# Stop if same tool called 5+ times, or A-B-A-B pattern detected
guard = FuzzyLoopGuard(max_tool_repeats=5, max_alternations=3, window=10)

Using with LangChain

If you're on LangChain, the callback handler does this automatically:

from agentguard import LoopGuard
from agentguard.integrations.langchain import AgentGuardCallbackHandler

handler = AgentGuardCallbackHandler(
    loop_guard=LoopGuard(max_repeats=3, window=6),
)

agent_executor.invoke(
    {"input": "Research AI safety"},
    config={"callbacks": [handler]},
)

pip install agentguard47

Zero dependencies, MIT licensed: https://github.com/bmdhodl/agent47

---

What loop patterns have you seen in your agents? I've seen exact repeats, A-B-A-B alternation, and slow drift. Curious what others encounter.