Agent loop with no iteration limit

An agent or tool loop with no iteration limit.

Rule agent-loop-unboundedCost audit

Why it costs money

An agent loop keeps calling the model until the model says it is done. When a tool keeps failing, or the model keeps asking for one more search, nothing in the loop itself ends the run. Each iteration is a full paid call, and in most agents each call also carries the growing transcript.

A loop like this does not fail loudly. It runs until the process is killed, a timeout hits, or the account runs out of credit. A bound that exists but defaults to None or -1 has the same effect on every caller who does not override it.

Example

import anthropic

client = anthropic.Anthropic()


def run_agent(task: str, tools: list) -> None:
    messages = [{"role": "user", "content": task}]
    while True:
        response = client.messages.create(
            model="claude-sonnet-4-5",
            max_tokens=2048,
            tools=tools,
            messages=messages,
        )
        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": execute_tools(response.content)})

import anthropic

client = anthropic.Anthropic()
MAX_STEPS = 15


def run_agent(task: str, tools: list):
    messages = [{"role": "user", "content": task}]
    for _ in range(MAX_STEPS):
        response = client.messages.create(
            model="claude-sonnet-4-5",
            max_tokens=2048,
            tools=tools,
            messages=messages,
        )
        if response.stop_reason != "tool_use":
            return response
        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": execute_tools(response.content)})
    raise RuntimeError(f"Agent did not finish within {MAX_STEPS} steps")

How to fix it

Give every agent loop a concrete maximum number of steps and an explicit stop condition, such as the model no longer requesting a tool. Raise or return a clear error when the limit is reached, so the caller knows the task did not finish.

If a limit is configurable, default it to a number rather than None or -1, and pair it with a token or dollar budget so a long run is stopped even when each step looks reasonable.

What the audit detects

• while True or while (true) loops whose body calls an LLM or agent and contains no break, return, raise or throw, exit, iteration counter or comparison.
• Agent limits such as max_iterations, max_turns or recursion limits whose default value is None, -1 or infinity, found by reading the default in the syntax tree, including typed parameters like max_iter: Optional[int] = None.
• Not flagged: loops that block on a queue or channel read (each iteration is triggered by an incoming message), recursion that counts down to a base case, limits set to a number, and optional parameters with no explicit default. Exits inside called functions are not traced.

Research

When Agents Do Not Stop: Uncovering Infinite Agentic Loops in LLM Agents. Static analysis detected infinite agentic loops with 91.9% precision on the paper's evaluation set, and treats a limit that defaults to a sentinel such as None or -1 as no limit.