Loop Guards: max_steps & Repeated-Action Halt

An agent that can loop can loop forever. The model gets confused, repeats the same failing tool call, and you wake up to a five-figure bill. Guard rails are not optional - they are the difference between a demo and something you would put in front of a customer. Two guards catch almost everything: a hard step ceiling, and repeated-action detection.

You will build run_with_guard(plan, tools, max_steps). It runs the same kind of plan as before, but instead of just a trace it returns a small report: {"status": ..., "trace": [...]}, where status is one of three outcomes:

• "max_steps" - the agent hit the step ceiling before finishing. Stop and flag it.
• "loop_detected" - the agent tried the exact same action (same tool, same args) it has already taken. That is a loop; halt before it spins.
• "done" - the plan completed normally (it ran out of steps or hit a "finish" action).

Track which actions you have seen so repeats are cheap to detect. A dict is not hashable, so turn each action into a hashable key first:

seen = set()
...
key = (action, repr(args))   # repr makes the args dict hashable
if key in seen:
    return {"status": "loop_detected", "trace": trace}
seen.add(key)

The trace truncates at the halt point: when you stop on max_steps or loop_detected, return the trace of what actually ran up to that moment, not the whole plan. That truncated trace is your incident report - it shows exactly where the agent went off the rails.

Check the step ceiling and the repeat before recording a step, so the counts line up: with max_steps=3 the trace holds at most 3 entries, and a repeated action is caught before it runs a second time. Press Run to watch all three outcomes fire.