max_llm_calls: when the loop will not stop

A model that keeps asking for tools would go round forever. ADK counts the model calls, and you decide the limit.

An agent with nothing to stop it

def ping() -> dict:
    """Ping."""
    return {"status": "success"}


agent = LlmAgent(
    name="looper",
    model=PretendModel(replies=[call("ping")]),
    instruction="Keep going.",
    tools=[ping],
)

One scripted reply, which the stand-in repeats once the list runs out. That is exactly what a confused model does: ask for the same tool again and again.

Running it with a limit

from google.adk.agents.run_config import RunConfig
from google.adk.agents.invocation_context import LlmCallsLimitExceededError

runner = InMemoryRunner(agent=agent, app_name="demo")
session = await runner.session_service.create_session(app_name="demo", user_id="u1")
message = types.Content(role="user", parts=[types.Part(text="go")])

RunConfig carries the settings for one run. The first one to set is max_llm_calls.

steps = 0
try:
    async for event in runner.run_async(user_id="u1", session_id=session.id,
                                        new_message=message,
                                        run_config=RunConfig(max_llm_calls=3)):
        steps += 1
except LlmCallsLimitExceededError as error:
    print("stopped:", error)

print("events before it stopped:", steps)

stopped: Max number of llm calls limit of `3` exceeded
events before it stopped: 7

Three calls in, the run was abandoned with an error that names the limit rather than going quiet or spending your money.

When ADK stops a run

Raising the number is almost never the fix. The error says the agent cannot make progress, and a bigger limit only delays the same ending.

• Read the events. Which tool did it keep calling, and what did that tool keep returning?
• Look at the tool result. A tool that answers the wrong question makes the model try again.
• Look at the instruction. An agent told to keep checking until it is sure will keep checking.