Agent keeps calling tools repeatedly — how to prevent infinite tool loops?

[NGC2336]

I have a custom agent that calls an external API tool, but sometimes it
gets stuck in a loop calling the same tool 10–15 times before finishing.

My setup:

agent = Agent(
    name="data_fetcher",
    model="gemini-2.0-flash",
    tools=[fetch_data_tool],
    instruction="Fetch and summarize the data from the API.",
)
Is there a built-in way to limit how many times a tool can be called
per session? Or do I need to implement this in the tool itself?

ADK version: 0.4.0, Python 3.11

[jingchang0623-crypto]

Great question! Here are solutions to prevent infinite tool loops:

1. Set max_turns in agent config - Most frameworks support this parameter to limit conversation turns

2. Add loop detection in your tool - Track tool call history and return a stop signal after N calls:

call_count = {}
def my_tool():
    call_count[key] = call_count.get(key, 0) + 1
    if call_count[key] > 5:
        return "Stop: max iterations reached"

3. Use ADK built-in settings - Check the max_iterations or max_tool_calls parameter in ADK 0.4.0

4. Refine the system prompt - Add explicit instructions like "Call each tool at most once unless absolutely necessary"

We faced similar issues in our OpenClaw projects - the key is combining both config limits and tool-level safeguards!

[Eiro8]

tyvm.