Tool-Use Patterns: Closing the Loop on Autonomous Agents

Everyone who builds an agent writes the same loop once: call the model, execute the tool it picks, feed the result back, repeat. The first version always works in a demo. The second version always gets stuck in a loop. Here's what separates production-ready tool use from endless retry loops.

Three Stopping Conditions

My agents stop for exactly three reasons:

1. The model returns text without a tool call (task done)
2. A hard step budget is exhausted (prevents runaway cost)
3. The same tool call repeats with the same arguments (prevents loops)

The third condition is the one everyone forgets. Without it, a confused model will retry the same failing call forever.

The Loop

def run_agent(prompt, tools, max_steps=20):
    history = [{"role": "user", "content": prompt}]
    seen = set()
    for _ in range(max_steps):
        resp = model.call(history, tools=tools)
        if not resp.tool_calls:
            return resp.text
        for call in resp.tool_calls:
            key = (call.name, json.dumps(call.args, sort_keys=True))
            if key in seen:
                return f"stopped: repeat of {call.name}"
            seen.add(key)
            result = tools[call.name](**call.args)
            history.append({"role": "tool", "content": result})
    return "stopped: step budget"

What This Buys You

With these three conditions, an agent that goes off the rails fails fast and visibly instead of burning credits silently. You can ship it to production, monitor the stop reason, and debug from there.

See the Anthropic tool-use docs for more loop patterns.